Cryptopolitan

According to Goldman Sachs, hedge funds suffered significant losses as the AI rally lost momentum in July. The bank said the pullback in AI-related stocks forced managers to unwind some of their heavy positions, resulting in one of the strongest de-grossing periods of the past 10 years.
It noted, “Our Hedge Fund VIP list of the most popular long positions suffered its worst 1-month underperformance vs. the S&P 500 in more than 20 years of history, and July marked one of the sharpest hedge fund de-grossing episodes of the past decade.”
At the moment, hedge funds are pulling back fast from AI stocks, according to the bank.
Hedge funds lost over 3% of their profits in July
Goldman asserted that hedge fund performance, leverage, and key long positions have shifted considerably as AI trade changed course. Data from across Wall Street also support this cooling-off period.
Similarly, JPMorgan in early August contended that tech sell-offs wiped out 3% of hedge fund gains in July.
According to their analysts, fund managers got trapped in overcrowded tech positions, creating a bottleneck that prevented speculators from cashing out before their profits vanished.
However, this summer slump might actually be part of a predictable seasonal pattern. JPMorgan noted that since 2018, hedge funds have tended to dump unprofitable stock positions in July. Because of this cycle, the bank hinted that traders could very well pick up tech stocks again by September, noting that managers frequently drop trades in mid-summer, only to buy back into the market in the coming months.
This year, when AI trade started losing momentum, analysts were still optimistic about AI trade and hedge fund performance. In late July, Vincent Lin, co-head of Prime Insights and Analytics in Global Banking & Markets, even noted that hedge funds were still deeply committed to AI tech.
At the time, he explained that the historic wave of tech selling looked more like a healthy market correction amid high volatility than a decline in confidence in AI. However, with traders currently moving away from AI, it’s unclear whether investors are still bullish on tech stocks.
Earlier this year, the war in Iran triggered a rough March for hedge funds. Though the funds rebounded quickly thanks to a massive chip stock rally led by Samsung, AMD, and SK Hynix.
The AI boom significantly contributed to the overly positive hedge fund performance in Q2
Primarily, the AI stock frenzy boosted second-quarter hedge fund performance, propelling investor crowding to historic heights. Per Goldman, tech stocks grabbed 14 out of 20 spots among the fastest-growing favorites on Wall Street.
Overall, according to data provider HFR, strong investment performance helped boost total industry assets by $409 billion, bringing the grand total to $5.6 trillion in the quarter. It also showed that macro strategies, where hedge funds make investment bets tied to indicators such as growth and inflation, took the crown as the most sought-after hedge fund style this year.
Speaking on the great performance back then, Shenan Dhanani, co-chief executive at Trium Capital, noted that this could be a “golden era” for the funds.
However, hedge fund performance has since slipped from those highs, though the funds are still outpacing their usual averages.
The concentration of hedge fund portfolios in AI-linked companies also made the July reversal more painful. Stocks connected to semiconductors, cloud computing, and AI infrastructure had attracted significant institutional demand during the rally, leaving many managers exposed to the same group of trades.
When momentum weakened, crowded positioning amplified losses as investors rushed to reduce their exposure simultaneously. This suggests that the July sell-off was not necessarily a rejection of artificial intelligence as an investment theme, but rather a warning that valuations and positioning had become stretched.
“Despite the volatility, US equity long/short hedge funds have returned 10% through mid-August,” Goldman said.
If hedge funds return to technology stocks in September, the latest pullback could prove to be little more than a summer repositioning.
However, continued weakness in AI-related shares could force managers to reassess the positions that helped drive their strong gains earlier this year.
Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.

Michael Saylor’s Strategy (NASDAQ: MSTR) is back to soaring with the eagles as Bitcoin’s climb past $77,000 put the firm back in profit on its 840,447 Bitcoin stash.
The Bitcoin position that took $63.36 billion to build is now worth about $65.5 billion, putting Strategy up more than $2 billion as of this Cryptopolitan report.
Strategy’s Bitcoin is now worth more than its cost basis. Source: Bitcoin Treasuries
That rally has trickled down to MSTR shareholders, with the stock up more than 11% in pre-market trading, rising from its $112.339 close on Thursday to about $126 early Friday.
Saylor and Strategy are back on top
Bitcoin’s march past the $77,000 mark has put Strategy back in the black. The firm paid an average of about $75,400 to build its 840,447 BTC portfolio.
CoinMarketCap data showed Bitcoin up nearly 10% in the last 24 hours, and more than 20% since its climb from the $63,000-$65,000 range started.
This BTC cycle started after two catalysts aligned for the Bitcoin market. First, the US Treasury said it would more than double its long-dated bond buybacks from $2 billion to $4 billion. That decision tipped the scales toward debasement hedges such as Bitcoin after the dollar lost its shine to weakness.
The sentiment snowballed into an avalanche of positivity after President Donald Trump hosted crypto executives, digital finance stakeholders, SEC Chair Paul Atkins and CFTC Chair Mike Selig at a White House gathering that rounded up to a Clarity Act push.
Buyers have continued to line up since. SoSoValue counted more than $783 million in bearish positions wiped out in 24 hours, part of roughly $3 billion in short liquidations.
Saylor’s firm stopped selling and started hoarding cash
This rally could not have come at a better time for Strategy after Saylor and other executives have had to come out to reiterate their long-term accumulation playbook after recent selling drew criticism.
The company did not sell Bitcoin during the week that ended August 16, the first in three weeks that it paused sales, as Cryptopolitan reported. Strategy last bought Bitcoin in June.
Instead, the firm has had to defend its books through dividend payments and preferred stock buybacks, while building up a $4.8 billion cash reserve.
The smartest crypto minds already read our newsletter. Want in? Join them.

Optimism token holders have voted to permit the transfer of 546.9 million OP from the collective’s user airdrop reserve.
The tokens are valued at ~$49 million. They land in a new pool the Optimism Foundation will manage. The voting closed on August 19, 2026.
Optimism voters clear the 51% threshold with a 62% total
The measure passed with 17,973,915 votes in favor against 10,930,696 opposed, the proposal page on Agora shows. That put support at about 62% of votes cast, above the 51% approval threshold.
The ballot also cleared its 16,540,389 quorum requirement. The proposal is now marked queued, with voting closed at 8:07 pm on August 19.
The vote came down to a late deciding vote from a team funded by Optimism itself. The onchain record only notes the running totals, without showing either the size of a single vote or the identity of the voter.
Filed by the Optimism Foundation, the proposal redesignates the OP left in the User Airdrop allocation as a Strategic Ecosystem Fund.
The document carries no substantive onchain transactions. It is a governance decision about how a large block of tokens gets labeled and who directs it.
The Foundation says that the renamed fund would bankroll adoption of OP Mainnet and OP Enterprise.
The proposal text covers “partnership deals that bring chains, protocols, institutions, and infrastructure to the OP Stack.”
It also contains incentives meant to strengthen activity and liquidity on OP Mainnet and plans to boost the OP Stack’s reach with institutions and large brands.
Tokens once earmarked for distribution to users become a discretionary pot for business development.
Optimism Foundation strengthens its grip on OP
As reported by Cryptopolitan at the time, the Collective passed OP-0017 in January, a proposal that would give half of the Superchain’s sequencer revenue to the Foundation to buy back OP each month.
The measure passed with 84.4%. The 12-month pilot kicked off in February, converting sequencer ETH into OP through an over-the-counter provider.
OP hit a low price of $0.2519 in December 2025. Since then, the Foundation has submitted several proposals that increase its power over the use of OP and treasury assets.
OP hit its all-time low of $0.08069 just three days ago, and even after this week’s bounce, the token sits only 25% above that floor.
It currently trades at $0.1008, up 10.7% on the day and 14% on the week, according to CoinGeko. The layer-2 platform token has a market cap of about $230.8 million and $57.8 million in 24-hour volume.
Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.

A syndicate recruiter has had his 56-month prison sentence confirmed by Hong Kong’s Court of Appeal in a fraud economy that INTERPOL estimates to have resulted in global losses of $442 billion in 2025. The flow of those illegal funds is being increasingly shifted to cryptocurrency exchanges and stablecoins.
The case around crypto reveals a larger issue. Chainalysis reported that no less than $14 billion was sent to fraud-related crypto wallets last year, with the number likely to amount to over $17 billion as scam wallets are uncovered. Stolen money is usually converted and laundered through the technology used by honest customers, which turns laundering related to scams into a very serious compliance headache for the bodies controlling crypto exchanges.
A guilty plea that put trafficking on the sentencing scale
According to reports, the appellant is Ma Che-hou, aged 32, who confessed to being involved in a conspiracy to defraud and money laundering in 2021 and 2022. The prosecutors stated that he convinced five men aged 20 to 32 by suggesting jobs with good pay, business chances, or online romance. Eventually, the men ended up in Southeast Asia, and some of them ended up in captivity in KK Park of Myanmar and were tortured, including electric shocks.
The case known as HKSAR v. Ma Che Hou [2026] HKCA 1479 involved an important legal loophole. As there is no distinct crime of human trafficking in Hong Kong, in this case, the judge saw the acts of trafficking and forced labor as relevant aggravating factors in the charge of fraud.
The court applied a base sentence of seven years and lowered it by one-third for Ma’s guilty plea, resulting in a sentence of four years and eight months. Judges noted that it was a good thing that the District Court’s seven-year maximum sentence limited the penalty; they said the crime was serious enough to merit a much higher maximum sentence.
USDT is the rail the money runs on
The connection between a prosecution in Hong Kong and the global cryptocurrency markets comes down to the infrastructure used to move the money.
According to the UNODC report for the year 2026, the criminal syndicates in Southeast Asia operate in a connected network where laundering, trafficking, and fraud operate independently but leverage the same equipment. The majority of the criminal profits are laundered using blockchain networks.
Delphine Schantz, UNODC Regional Representative for Southeast Asia and the Pacific, described the model this way:
“Their operating model looks like corporate franchising: imagine specialised departments for laundering money, trafficking people, smuggling migrants, and harvesting data.”
According to Chainalysis, there was an increase of 85% in the flows of cryptocurrencies towards fraudulent human trafficking services in 2025 when compared to the previous year. Stablecoins are preferred for payments because they are able to preserve value and can be easily converted into local currencies through money laundering networks that operate in China.
The public blockchain also provides an opportunity for investigators that cash does not: transactions leave clues for them. INTERPOL mentioned that there was a 20-year-old suspect located in Thailand who made over $122.5 million in romance scam transactions, from cross-chain swipes meant to hide their sources, over the time span of 10 months.
Seizures now run into the billions
Enforcement actions have now reached a significant number. According to Chainalysis, the Scam Center Strike Force of the U.S. Department of Justice (DOJ) announced in April 2026 that it had seized around $701.9 million in cryptocurrency found to be connected to money laundering activities and taken down a total of 503 fake investment websites. In addition, OFAC placed sanctions on 29 Cambodia-related persons and organizations, including Senator Kok An.
In another important case, the head of the Prince Group, Chen Zhi, was indicted by the DOJ and an extensive seizure of Bitcoin (around 15 billion dollars) took place. The U.S.-China Economic and Security Review Commission describes the seizure as the largest one in history.
A move of such magnitude decreases the liquidity available to criminal networks and indicates to the exchanges that dealing with illicit funds – knowingly or not – becomes more and more legally risky.
FATF puts fraud at the center
The regulatory path is now beginning to become clearer. On July 1, 2026, Financial Action Task Force (FATF) President Giles Thomson made the occasion of his first day in office a significant one with the launch of a multi-year roadmap addressing fraud as a priority. FATF estimated nearly $500 billion in total global losses due to fraud during 2024-2025. Furthermore, FATF reported that nearly 90% of the assessments in the last round of mutual evaluation have indicated fraud as a key crime that generates proceeds. Under the roadmap, it will be analyzed how countries may improve their responses to fraud as well as the money laundering associated with it, with the policy recommendations expected to come in by February 2027.
Thomson summed up the urgency:
“Fraudsters and other criminals are scaling at speed by exploiting technological innovations, often targeting the most vulnerable in society.”
For crypto companies, it means stricter controls with regard to transactions, particularly with mule accounts and fast cross-border transactions. UNODC has also called for specialized training for regional law enforcement so officials can trace, identify, seize and recover criminal proceeds moving through crypto. This shows the growing recognition that arresting ringleaders alone has not been enough to slow the crypto-fraud industry.

The smartest crypto minds already read our newsletter. Want in? Join them.

OpenAI activated a feature on Thursday that hands ChatGPT the keys to Apple Messages on the Mac.
The AI assistant can read, search, draft, and send iMessage, SMS, and RCS texts on behalf of the user.
ChatGPT reads, drafts, and sends texts inside Messages
When connected, ChatGPT loads a user’s Messages inbox to sort, analyze, and edit texts.
It can dig up old information and draft replies or delete messages on command. ChatGPT can also summarize conversations and answer questions about them.
A promotional clip shows someone asking the assistant to propose follow-ups based on texts that received the day before.
The plug-in is also integrated with Codex and ChatGPT Work, so the same message handling reaches into a professional account.
OpenAI advises users to watch what the assistant is doing and warns against switching on persistent approval.
The setting “removes your final chance to review a message before ChatGPT sends it as you,” the company said via its website.
OpenAI stays vague on data as Apple ties fray
Apple has kept Messages closed off as a key part of its brand.
OpenAI said the plug-in runs locally on the Mac and “doesn’t create an index of all someone’s messages.” The AI company did not clarify what that covers.
In June of 2026, as Cryptopolitan reported, OpenAI rolled out Lockdown Mode across ChatGPT tiers. The optional setting cuts off web browsing and agent capabilities to reduce prompt-injection attacks that may expose sensitive information.
According to a May report, OpenAI had hired outside attorneys to look into a lawsuit against Apple over their Siri-ChatGPT deal, which the AI company blamed for not bringing in the paying users it had expected.
The plug-in builds ChatGPT control directly into Apple’s own Messages app. In January 2026, Apple dropped OpenAI in favor of Google’s Gemini.
Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.

MANTRA Chain temporarily halted network operations on Thursday as a precautionary measure following an undisclosed incident. It is freezing all public endpoints and on-chain transactions. As all the validators are currently paused, all transfers of its native token, OM, are completely disabled.
Consequently, major South Korean and international cryptocurrency exchanges have suspended OM deposits and withdrawals. While spot trading remains active on these centralized platforms, users cannot move their assets on-chain or withdraw funds until the core team and security partners conclude their investigation and safely restart the network.
For a sector that invested much of 2025 mulling over whether real-world asset tokens had been mistakenly vilified, the outage brings about an unwelcome query into the durability of OM’s infrastructure. Now, the market cap of the token is currently pegged at roughly $25 million after the losses of last year, meaning the financial impact is quite insignificant. However, the reputation loss may be quite tricky to handle as the OM is the same asset that plummeted by over 98% in one April session in 2025, causing a dip in the confidence of many people in the RWA market.
A precautionary freeze with no clock on it
According to the status page of MANTRA, their team of engineers and security specialists has stopped the chain after realizing there has been an incident, and they are looking into it together with partners. The chain cannot restart until the team is sure it is safe to do so, and until then, no transactions can be executed and OM balances are frozen in place.
The team has also alerted users about a common secondary risk during downtime, namely that of scammers. They told holders that they should use only official channels and avoid anyone reaching out to them offering to help retrieve their funds.
What is behind the events still remains a mystery. The company has not indicated what the reason behind the incident is, whether any sort of hack has taken place, or whether user funds are at risk.
Upbit and Bithumb pull OM off the rails
The fallout reached South Korea quickly, where OM has an active trading base. Upbit, one of the country’s largest exchanges, suspended OM deposits and withdrawals because of the network issue, according to Bitcoinworld. Bithumb followed with its own suspension for the same reason.
According to MANTRA, it has made the relevant notifications to exchanges and ecosystem partners; other than that, deposits and withdrawals on the impacted exchanges have been suspended until the chain restarts trading.
But spot trading may carry on. According to Bitcoinworld, users are still able to buy and sell OM via Upbit, even though transfers have been disabled. That makes for an interesting market, where prices can fluctuate, but traders cannot add new OM or withdraw the one they have.
A live order book that is out of sync with unusable settlement channels creates less liquidity in the market, making it more susceptible to major price fluctuations. The situation reminds one of the situation prior to the last significant crash of OM.
Why a small token still worries the RWA trade
OM is now a fraction of its former size, which makes the timing especially awkward for MANTRA. The chain’s transparency report put OM at about $0.0054 and a $29.9 million market cap in early August, with 5.53 billion tokens circulating. At its 2024 peak, the token was worth as much as $6 billion, according to earlier Cryptopolitan reporting.
After the April 2025 crash, CEO John Patrick Mullin said: “We have determined that the OM market movements were triggered by reckless forced closures initiated by centralized exchanges on OM account holders.”
Since that time, MANTRA has been working on the restoration of its institutional position. In June, Inveniam Capital Partners agreed to purchase MANTRA and its associated companies, with the deal anticipated to be completed in the third quarter of 2026 and MANTRA is also in possession of the digital-asset license from Dubai’s VARA.
An unexplained chain halt cuts directly against that recovery story. For the wider RWA market, the episode is a reminder that regulatory positioning and institutional backing do not eliminate infrastructure risk. A compliance-focused Layer 1 can still go dark without warning, and OM’s comeback still depends on a network investors have already seen tested before.

Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.

Japan’s S&P Global Flash Manufacturing PMI rose to 55.1 in August, up from 54.5 the previous month. New orders surged at their fastest rate since January 2018. This growth was driven by major international AI clients, including foundries, chipmakers, and data center operators.
For producers of components and equipment in Japan, the report suggests that investment in AI is continuing to promote industrial activity in the country.
Where are the orders coming from
The July regional survey of the Bank of Japan, released on the 9th, provides a clearer understanding of what is pushing the demand. According to branch managers, orders for equipment used in semiconductor production and electronic components have increased with the ongoing growth of global AI investment. At the same time, demand starts to move into other sectors such as equipment for electrical power generation, communications equipment, and industrial molds.
Hard production data confirms the trend. According to a revision by the Ministry of Economy, Trade and Industry issued on August 17, Japan’s industrial production index improved to 104.6 in June, seasonally adjusted data reflecting an increase of 1.9% compared to May and 4.9% year-on-year. The factory’s operating ratio, which is an indicator of how much factory work went up by 4.1% since the previous month.
The export side of the same story
The recent trade data from Japan provides a similar outlook. Exports grew by 23.2% year-on-year in July to an unprecedented ¥11.5 trillion, exceeding projections of a 19.9% increase and growing faster than a 19.3% rise in June, according to a Reuters report. The rise was bolstered by demand related to artificial intelligence data centers, while a weak yen contributed to Japanese goods being more competitive in global markets.
Shipments to the U.S. increased by 22%, and exports to China increased by 25.8%. On the other hand, imports grew by 27.8% to a historically high figure of ¥12.1 trillion, mainly due to rising oil prices, thereby leaving Japan with a trade gap of ¥634.5 billion.
What it signals for the wider AI buildout
The strength showing up in Japanese factories is consistent with broader industry forecasts. SEMI expects worldwide spending on 300mm fab equipment to rise 18% to $133 billion in 2026 and another 14% to $151 billion in 2027, crossing $150 billion for the first time.
Japan is not only supplying the AI buildout; it is investing heavily in it at home. The trade ministry said in July it would purchase 27,500 of Nvidia’s (NASDAQ: NVDA) next-generation Rubin chips for Noetra, a SoftBank-led sovereign AI project backed by ¥1 trillion over five years, as Cryptopolitan reported.
Bringing all the information together, we can conclude that Japan is benefiting from AI investments at all stages of the supply chain due to increased factory orders, record levels of exports and domestic purchases of chips.
There is one significant restriction in this situation. As the BOJ survey indicated, small companies are still facing challenges in transferring their increasing input costs to customers, while producer prices are up 7.2 percent year-on-year in July. Due to strong exports, this inflationary pressure could justify another rise in interest rates by the Bank of Japan, which might take place as early as September.

Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.

Sentiment across the crypto market has flipped bullish over the past three days. Roughly $313 billion has been added to the crypto market, with the total market cap rising from $2.19 to $2.5 trillion. Wednesday, August 19, saw the biggest short liquidation in crypto history with over $2.7 billion worth of positions being wiped out, according to CoinGlass. As Crypto turned higher on Wednesday, Kalshi set two new records on the same day.
Data from Artemis shows that the platform’s daily crypto spot volume reached a new high of $268.89 million. Meanwhile, daily perp open interest hit $18.69 million. Both these prints landed just as the market rallied but they need to be read separately and they track different behaviour.
Source: Artemis
What Kalshi’s Spot Volume Actually Counts
Artemis labels the crypto-based event contracts as spot volume on the page. Kalshi does not run a spot order book and these are basically the yes or no binaries based on whether a specific asset closes above a given level by a set date. A trader’s downside and risk is based on how much he put down for a contract and there is no liquidation. This design draws flows in regardless of which direction the market is moving, so volume records on these books measure attention rather than conviction.
In early June, we covered how Kalshi printed a new crypto volume record at the time and how it took place during the heaviest liquidation day since February. Therefore, the $268.89 million number right now shows that crypto event-based contracts is the place where traders go when the market moves and it says nothing about directionality.
Open Interest Is Where the Direction Shows Up
Crypto perpetual futures launched on Kalshi on June 3 and volume cleared the $1 billion within the first week. Open interest is a valuable metric because it shows what stayed on the table after the close. A record OI reading on the day a rally began means traders were carrying leveraged exposure overnight instead of scalping in and out during the session.
That is the number pointing somewhere. Positions held through the close are a bet on continuation, and the Artemis chart shows the metric climbing from roughly $2 million on June 4 to $18.69 million on August 19, with only a flat stretch through late June breaking the trend.
Small Absolute Numbers, Steep Trajectory
Nobody is claiming $18.69 million in open interest competes with Binance or Hyperliquid, where perp OI runs into the billions. On size alone, Kalshi barely registers in the perp landscape.
The venue type is what makes it worth tracking. Kalshi holds a CFTC license and operates onshore, so leveraged crypto demand that historically routed to offshore exchanges is now clearing through a US-regulated market. Eleven weeks of data is a short sample, but the direction of the line has been consistent since launch.
Prediction market platforms have spent the past year moving from election contracts into financial markets. Perps put Kalshi in direct competition with crypto-native venues rather than adjacent to them.
The Data Already Lags the Tape
One caveat on timing. These are the most recent figures Artemis has published, and crypto has continued to climb since August 19.
Both records may have been taken out already. The next Artemis update will show whether the August 19 highs held for more than a day, and whether the open interest build survived the move higher or got unwound into strength.
If you're reading this, you’re already ahead. Stay there with our newsletter.

Broadcom is negotiating to secure over $60 billion, possibly up to $100 billion, to support the production of AI chips for Anthropic and its other clients. This transaction highlights the reality of how the development of AI across the world is being financed by loans rather than company funds.
Debt, not cash, is bankrolling the chips
The financing being discussed will be divided into various portions. One of these includes the junior debt portion estimated at about $30 billion. According to reports, Broadcom is also expected to guarantee part of its senior secured debt portion worth approximately $60 billion to $70 billion. The total amount of financing that will be raised is expected to reach $100 billion.
A special-purpose vehicle would issue the debt, keeping it off Broadcom’s own balance sheet. Blackstone and Apollo Global Management, two of Wall Street’s biggest private-credit firms, are also in talks to participate.
This is not a one-off transaction. Goldman Sachs Research predicts total debt issuance associated with AI could reach just under $500 billion by 2026. Credit strategist Amanda Lynam put it succinctly:
“It’s hard to overstate the importance of this theme in the credit markets, both in terms of its overall scale.”
The general conclusion is evident. The cost of developing AI infrastructure has reached a level at which even companies with enormous cash reserves do not want to cover the costs fully by themselves.
A financing model built to loosen Nvidia’s grip
What makes the financing strategically important is what Broadcom’s chips are designed to do. The company develops custom silicon for Alphabet and Meta and has supply agreements with Anthropic and OpenAI, as major AI players look to build their own accelerators and reduce their dependence on Nvidia.
Providing funding for such custom chips at this magnitude makes that option more achievable.
Nvidia is working towards a similar end from the other direction. The chip manufacturer reportedly announced in August that it had arranged funding with investment companies like Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs, and KKR. This was done in order to obtain over $500 billion in third-party financing.
That changes the nature of the AI chip race. Competition is increasingly about who can secure the cheapest and deepest pools of capital, not simply who can build the fastest processor.
The 20-gigawatt bet behind the numbers
The new raise builds on a model Broadcom established in June, when it, Apollo, and Blackstone launched a platform with an initial $35 billion transaction to expand Anthropic’s computing capacity by more than one gigawatt.
Ultimately, this collaboration’s objective is to bring 20+ gigawatts of compute capacity to frontier AI laboratories such as Anthropic and OpenAI by 2028. This news indicates that the new debt agreement might have the same shape as the previous ones.
The figures might become significantly bigger. According to Bank of America’s Tom Curcuruto, Broadcom’s chip financing facility could advance to $370 billion worth of senior debt by the middle of 2029, which would be used to finance 20 gigawatts of capacity.
The electricity requirement gives the scale some perspective. As Cryptopolitan mentioned in its April report, a one-gigawatt data center needs approximately the same amount of energy as one million homes in the US.
Broadcom’s AI revenue is already climbing fast
The borrowing comes as Broadcom’s AI business accelerates sharply. In its fiscal second quarter ended May 3, 2026, the company generated $10.8 billion in AI semiconductor revenue, up 143% from a year earlier.
CEO Hock Tan told investors he expects that figure to exceed $16 billion in the third quarter, representing growth of more than 200%. Total quarterly revenue reached $22.2 billion.
That growth helps in understanding the willingness of lenders to fund hardware on such a large scale. Broadcom is not merely speculating on future demand for AI since its current revenue already reveals how fast the demand is turning into chip sales.

Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.

If you search for a robot dog in 2026, you are likely to get lost in comparing backflipping quadrupeds patrolling factories and a $319 kit that fits in your hand. Since robodogs look similar, it is not uncommon for them to be grouped. However, such a comparison will lead to the wrong decision.
Some robots are just interactive pets, while others may be research-based robots designed for developers and research teams. They serve different purposes. For instance, parents may be more concerned about ease of use and need a fun pet robot for their kid. A developer or a research student, however, would need more programmability. Similarly, industrial buyers care more about the payload, endurance, and autonomous behavior of the robot for specific tasks. There’s also a separate category of industrial machines.
Hence, the buying decisions are completely different for each category.
This guide covers a range of robodogs with prices ranging from $319 for Petoi Bittle X to $100,000 for Unitree’s industrial platform. We have separated the categories into three tiers to guide the buyer to the right tier and pick the best quadruped robot dog as per their needs. Readers looking for bipedal systems can use our separate guide to humanoid robots.
Quick Comparison
Model Tier Price (source + date) Programmability Runtime Support reality Best for Petoi Bittle X Companion / Coding From $319 (Petoi, Aug. 9) Blocks, C++, Python, ROS ~1h walking Direct; 1yr warranty; parts available Robotics learning and coding KEYi Loona Companion / Coding $499 sale / $529 regular (KEYi, Aug. 9) Blockly app; no official SDK Up to 2h claimed; ~60–90m active Direct; 1yr warranty; cloud-linked features Interactive robot pet Sony aibo ERS-1000 Companion / Coding $3,199.99 incl. 3-year cloud plan (Sony US, Aug. 9) Visual Programming, Web API ~2h Sony repair/parts; cloud-dependent Premium robot companion Unitree Go2 Developer / Prosumer Air $1,600/$1,850; Pro $2,800/$3,050; X $4,500; EDU quote (Unitree, Aug. 9) Air/Pro graphical; X partial dev; EDU SDK2/Python/ROS2 Air/Pro/X 1–2h; EDU 2–4h Direct; Air 6mo, Pro/X/EDU 12mo Affordable legged robotics Unitree Go1 Developer / Prosumer Air $2,700; Pro $3,500; EDU quote (Unitree, Aug. 9) EDU: C++, Python, ROS1 Not officially specified Direct; legacy platform; checkout unavailable Existing Go1 workflows DEEP Robotics Lite3 Developer / Prosumer Basic $2,890; Basic AI $4,968; Venture $6,480; Pro $12,510; LiDAR $18,270 (official US partner, Aug. 9) SDK/API, C++, Python, ROS1/ROS2 1.5–2h; Basic AI not specified Official US partner; 3–12mo component coverage ROS-ready research Unitree Aliengo Industrial / Inspection $50,000 displayed/reference (Unitree, Aug. 9) Legacy SDK, C++, Python, ROS1 2.5–4.6h Direct industrial; warranty term not public Industrial R&D Unitree B2 / B2-W / A2 Industrial / Inspection $100,000 displayed/reference each (Unitree, Aug. 9) C++, Python, ROS2 B2/A2 >5h unloaded; B2-W distance-rated Direct industrial; A2 12mo warranty Heavy-duty payload and inspection Boston Dynamics Spot Industrial / Inspection Quote only (Boston Dynamics, Aug. 9) Python/C++ SDK, gRPC, ROS2 ~90m; ~60m with powered payloads Enterprise; 1yr warranty + Spot Care Enterprise inspection
How to Choose a Robot Dog
One may think the robot dogs are all similar, even if they differ in their hardware and specifications. Before you begin searching for one, you should know why you want it and what purpose you want it to serve. That’s where it becomes important to know about the tiers and which one suits you.
The Three Tiers
Companion: As the name suggests, this tier is for people who are looking for a pet-like experience and entertainment. They don’t want to dig much into the technical setup. Such robot dogs mostly come pre-programmed with easy-to-use functions and handling. The buyers for such robot dogs are curious hobbyists, parents who want to keep their children busy with an interactive and learning-based toy, or elderly people who need someone to talk to.
Developer: This tier is suitable for tech geeks who are interested in the know-how of things. They want to modify it for their own needs and have coding and programming knowledge. Mostly university research students who write code, engineer gathering real-time sensory data are interested in them. In short, it’s for research purposes.
Industrial: These robots are made for specific tasks needed for a regular industrial job. They can be used for taking daily rounds at an oil refinery for any gas leaks, dangerous mine inspections where humans can’t go, and logging data for longer hours
What you’re paying for
The price jumps between the three tiers as they differ in the actuators, sensing, onboard compute, autonomy, and the amount of hardware the robot can carry.
A lower-cost Tier 1 robot has basic cameras, touch sensors, or basic obstacle detection for the interactive functions it is needed for. Developer platforms, Tier 2, add LiDAR, depth cameras, expansion ports, and more computing power.
Industrial models, Tier 3, go further with heavier-duty joints, larger batteries, environmental protection, payload interfaces, and sensors designed for inspection or navigation. Payload also matters. Smaller developer robots are generally built to carry a few kilograms at most, while Unitree’s B2 and B2-W are designed to move more than 40 kg. That difference changes both the hardware and the price.
Programmability: SDK, ROS support, and community size
Programmability is not limited to robot dogs with complex functions. Even the simplest companion robot dogs have some extent of programmability in them, which only requires tapping buttons in an app or dragging visual blocks around a screen without expert coding knowledge. Programmability preference increases with expertise, which requires writing full software to control every movement, sensor, and decision the robot makes.
Two things worth checking are SDK and ROS support. An SDK is basically a toolkit that comes along with the product for the developer to build their own programs for the robot. ROS is a widely used robotics software standard that lets developers use community-built tools rather than starting from scratch. If a robot supports both, it opens up far more possibilities.
Community size also matters here. An active community means more shared code, tutorials, and fixes already exist online, which makes development faster and less frustrating.
Cloud dependency and its risks
Some robot dogs can operate largely on local hardware, while others depend heavily on the manufacturer’s software and cloud services.
Bittle X can function without a network connection. Loona can still perform many basic interactions offline, but setup, voice recognition, firmware updates, and connected features rely on online services. aibo goes further. Its full experience is tied closely to Sony’s A.I. Cloud Plan and My aibo ecosystem.
Sony’s aibo shows why this matters in practice. Japan sales are ending, and Sony has continued repairs, parts, and cloud service for owners, but a robot whose personality lives on a subscription is only as durable as the vendor’s commitment to it.
Runtime, noise, terrain, and other things that demo videos hide
Demo videos are made to impress, not inform. Battery life is almost always shown at its best, However real use drains faster, and capacity quietly shrinks after months of regular charging. Noise is also easy to hide with background music, but motors hum every time the robot moves. Terrain looks effortless on camera if filmed on ideal surfaces, not rugs, thresholds, or slight slopes. Recharge time and long-term wear never make it into a two-minute product video either. These are the things worth checking before buying, because they are what actually shape the experience of owning the robot week after week.
Tier 1: Companion and Coding Robot Dogs ($319–$3,199.99)
1. Petoi Bittle X
Petoi Bittle X
Petoi’s Bittle X is the entry point for buyers who want to learn how a quadruped works rather than simply watch one perform tricks. It is a small nine-degree-of-freedom robot dog built around Petoi’s ESP32-based platform, and it is available either as a construction kit or pre-assembled. The construction version takes roughly 40 to 90 minutes to put together, which is part of the appeal for students and hobbyists but less useful for someone looking for a plug-and-play robotic pet.
Petoi’s direct store (as of Aug. 9, 2026), lists the Bittle X from $319 for the construction version with lite servos. A pre-assembled lite-servo version costs $339, while alloy-servo configurations cost $379 assembled as a kit or $399 pre-assembled. Petoi’s original Bittle is now marked Final Stock, so Bittle X is the more relevant model for a new buyer in 2026.
Programming is where Bittle X stands out. Beginners can start with block coding, while Python, C++, and ROS give more experienced users room to build deeper projects. It can also work without a network connection, and Petoi provides replacement parts, repair guidance, and support through email, chat, and remote video calls.
Runtime is modest at about one hour of continuous walking, with sustained running capable of cutting that below 30 minutes.
Pros
Low-cost entry into real quadruped programming
Broad path from block coding to Python, C++, and ROS
Repairable, expandable, and not dependent on cloud services
Cons
Limited battery life under active use
Best suited to indoor, relatively flat surfaces
The construction version requires assembly and setup
2. KEYi Tech Loona
KEYi Tech Loona
Loona takes a different approach from Bittle X. It is designed first as an interactive robot pet, with coding added as a secondary feature. It is also not truly a “legged” quadruped because Loona moves on four wheels.
It uses its body movements with ears, head, and wheels for actions to create much of its personality. The most expressive part is the eyes, which are incorporated in the 2.4-inch LCD on its head. The same screen is also used to show visuals and animations. Moreover, it has a camera, microphones, and touch sensors for better interaction. Loona’s voice conversations are powered by GPT-4o. According to KEYi, it gives its spoken interactions a noticeably more natural, open-ended feel.
KEYi’s direct store (checked Aug. 9, 2026) lists Loona at a $499 sale price, down from a regular price of $529. It is aimed mainly at families and buyers who want games, expressive interaction, voice features, and a robot that can move around the home without requiring much setup.
Programming is available through Google Blockly inside the Hello Loona app. That makes basic behavior programming approachable for children and beginners, but the ceiling is much lower than on Bittle X. KEYi currently provides no official standalone SDK or Python development environment.
Battery life also needs some context. KEYi advertises up to two hours of continuous play, but its own 2026 material puts more active use at roughly 60 to 90 minutes. Loona can return to its charging dock automatically.
Support includes a one-year warranty and advertised lifetime customer service. Many basic interactions work offline, but initial setup, voice recognition, firmware updates, and some connected features depend on internet services.
Pros
Stronger companion and entertainment experience than a coding-focused kit
Beginner-friendly Blockly programming
Automatic charging-dock behavior
Cons
No official standalone SDK for deeper development
Several important features depend on cloud services
Wheeled design limits its value for true quadruped locomotion or terrain work
3. Sony aibo ERS-1000
Sony aibo ERS-1000
Sony’s aibo is the premium companion option in this tier. At its launch in 2018, the sales reached 20,000 in the first six months. It is built around interaction rather than robotics development, using 22 movable axes, animated OLED eyes, cameras, microphones, touch sensors, and other onboard sensors to create more natural movement and responses around the home.
Sony’s US store (checked Aug. 9, 2026) lists aibo at $3,199.99, including a three-year A.I. Cloud Plan. That cloud connection is an important part of the product. aibo’s memories, learning, My aibo features, and parts of its evolving behavior depend on Sony continuing to maintain the service. aibo is still programmable, but in a much more controlled way than Bittle X. Sony provides Visual Programming for simpler behavior creation and a Web API for developers, rather than a low-level robotics or ROS environment.
Runtime is about two hours in standard operation, with roughly three hours needed for a full recharge. aibo can return to its charging station automatically when the battery runs low.
There is also a lifecycle point buyers should know. Sony ended ERS-1000 sales in Japan in June 2026, but US sales continue. However, Sony has also said it will keep providing technical support, replacement parts, repairs, cloud plans, and My aibo services for existing owners. Sony has not announced a successor model, so buyers are getting three years of committed cloud access with nothing published about what comes after.
Pros
Highly developed companion behavior and physical expressiveness
Automatic charging and mature home interaction
Continuing Sony repair, parts, and software support
Cons
Expensive at more than $3,000
Full experience depends heavily on Sony’s cloud ecosystem
Limited compared with open robotics platforms for deeper development
Tier 2: Developer and Prosumer Quadrupeds ($1,600–$18,270)
4. Unitree Go2
Unitree Go2
Unitree’s Go2 is where the category starts to look less like a robot pet and more like a compact robotics platform. Go2 made LiDAR-equipped, programmable robot dogs affordable for university-level research students under $2000. It is also a part of a much wider product expansion at Unitree, which has been moving toward an IPO, as reported earlier by Cryptopolitan.
The Go2 weighs about 15 kg and comes with 4D ultra-wide LiDAR and an HD camera across the current range, but what you can actually do with it depends heavily on the configuration.
Unitree’s direct store (checked Aug. 9, 2026), lists the Go2 Air from $1,600 without a controller or $1,850 with one. The Pro costs $2,800 without a controller or $3,050 with one, while the X is $4,500 with a controller. However, the EDU version is quote-only.
That price ladder also changes the development access. Air and Pro support graphical programming but do not provide Unitree’s secondary-development interface. X adds partial secondary development, while EDU is the version with full SDK2, Python, and ROS2 access. For developers, that distinction matters more than the headline starting price.
Runtime is one to two hours on Air, Pro, and X. EDU uses the larger 15,000mAh battery and is rated for two to four hours.
Support is direct from Unitree. Air carries a six-month warranty, while Pro, X, and EDU get 12 months. Buyers should also account for international shipping, customs, and the fact that Unitree does not offer standard returns on direct purchases.
Pros
Very low entry price for a LiDAR-equipped legged robot
Multiple configurations covering prosumers through research use
Full SDK2, Python, and ROS2 support on EDU
Cons
$1,600 Air does not include full developer access
Standard configurations still run for only one to two hours
China-direct buying can add shipping, customs, and support friction
5. Unitree Go1
Unitree Go1
The Go1 is Unitree’s previous-generation quadruped, and in 2026 its main appeal is for buyers who already have a reason to stay with the platform. It weighs about 12 kg and was built as a compact consumer and research robot, but the newer Go2 now starts at a lower price and uses Unitree’s more current development ecosystem.
Unitree’s direct listings (checked Aug. 9, 2026), show $2,700 for the Go1 Air and $3,500 for the Pro, while the EDU version is quote-only. However, the robot is currently unavailable through Unitree’s normal online checkout. Unitree has not announced that the Go1 is permanently discontinued, so it remains a currently listed product with uncertain direct availability rather than a formally retired model.
Programming access also depends on the version. Air and Pro do not include Unitree’s scientific or Python programming interfaces. The EDU model is the developer-focused option, with C++, Python, and ROS1 support through Unitree’s older software stack.
Unitree does not currently publish a verified hour-based runtime for Air, Pro, or EDU, so battery capacity alone should not be used to estimate it.
Support continues through Unitree, although warranty coverage is shortest on Air: six months for core components and three months for non-core parts. Pro gets 12 and six months, respectively, while EDU carries 12 months.
Pros
Established SDK and ROS1 ecosystem for EDU users
Compact 12 kg platform
Still supported and not officially discontinued
Cons
Currently unavailable through normal direct checkout
Older software stack than Go2
New buyers get a stronger price proposition from the newer Go2
6. DEEP Robotics Lite3
DEEP Robotics Lite3
The Lite3 is the main non-Unitree option in this tier. DEEP Robotics sells it as a research and secondary-development quadruped, with five versions that add more onboard compute, sensing, and navigation capability as the price rises.
DEEP Robotics’ official US partner store (checked Aug. 9, 2026) lists Lite3 Basic at $2,890. Basic AI is $4,968, Venture is $6,480, Pro is $12,510, and the LiDAR version is $18,270. That gives research teams a much wider configuration range than a single fixed platform.
Programming support is one of Lite3’s strongest points. The platform supports SDK and API development with C++, Python, ROS1, and ROS2. Unlike the lower-priced Go2 Air and Pro, developer access is central to the product rather than reserved for a top research configuration.
Runtime is 1.5 to 2 hours on Basic, Venture, Pro, and LiDAR. DEEP Robotics does not publish a separate runtime figure for Basic AI. Payload also falls as more hardware is added, from about 5 kg on Basic to 2.5 kg on LiDAR.
Support in the US comes through DEEP Robotics’ official partner channel, with technical support and after-sales service. Warranty coverage is shorter on wear components, and Basic AI’s joint warranty is only three months.
Pros
Strong C++, Python, ROS1, and ROS2 development support
Five configurations covering basic research through LiDAR navigation
Credible alternative for labs that do not want to standardize on Unitree
Cons
Higher configurations become expensive quickly
Payload drops as more sensing hardware is added
Short warranty coverage on joints and other wear components
Tier 3: Industrial and Inspection Platforms ($50,000+; Spot quote-only)
7. Unitree Aliengo
Unitree Aliengo
Aliengo sits between Unitree’s smaller developer robots and its newer heavy-duty industrial platforms. It weighs about 21.5 kg without the battery and is rated to carry around 13 kg, which gives it substantially more payload capacity than Go2 or Lite3 without moving into the size and weight of B2.
Unitree’s direct store (checked Aug. 9, 2026) displays Aliengo at $50,000, although that should be treated as a reference price rather than a normal retail checkout figure. The actual purchase process is sales-led.
The robot is built for industrial R&D and field work, with depth cameras, visual odometry, optional LiDAR, and expansion interfaces for equipment such as additional cameras, GPS, robotic arms, and other payloads. Runtime is also stronger than the smaller developer platforms, at 2.5 to 4.6 hours.
Aliengo is programmable through Unitree’s older legged-robot software stack, with C++, Python, and ROS1 support. That remains useful for existing projects, but it is not the same as the current SDK2/ROS2 environment used by newer Unitree platforms.
Support is handled directly through Unitree’s industrial and after-sales channels, although a current public Aliengo-specific warranty duration has not been verified.
Pros
13 kg payload is a clear step up from developer-tier robots
2.5 to 4.6 hours of endurance
Designed for external sensors and payload integration
Cons
Uses Unitree’s older SDK and ROS1 stack
$50,000 is only a displayed reference price
Current public warranty duration is unclear
8. Unitree B2 / B2-W / A2
Unitree B2 / B2-W / A2
Unitree’s B2, B2-W, and A2 move into a much heavier class of industrial robots. These are built around payload, endurance, and inspection work rather than portability.
Unitree’s direct store (checked Aug. 9, 2026) displays $100,000 for each platform, although the final purchase process is sales-led and the displayed figure should be treated as a reference price.
The B2 weighs about 60 kg and is rated to carry more than 40 kg while walking, with at least 120 kg of standing load capacity. It can operate for more than five hours unloaded and more than four hours with a 20 kg payload. The B2-W uses a wheel-legged design and weighs about 85 kg, combining rough-terrain movement with longer travel distances of around 30 km unloaded or 25 km carrying 40 kg.
The A2 is lighter at about 42 kg with its battery and is rated for around 25 kg of continuous walking payload. It can run for more than five hours unloaded or more than three hours with 25 kg, and its dual hot-swappable batteries are better suited to longer industrial deployments.
All three support secondary development with C++, Python, and ROS2 interfaces. Unitree also provides direct industrial and after-sales support. A2 has a confirmed 12-month warranty, while current public warranty durations for B2 and B2-W have not been verified.
Pros
Much higher payload capacity than developer-tier quadrupeds
Longer endurance for inspection and field work
ROS2-based development and industrial sensing support
Cons
$100,000 displayed pricing puts them firmly in enterprise territory
Large and heavy compared with developer robots
Exact capabilities and warranty terms vary by model and configuration
9. Boston Dynamics Spot
Boston Dynamics Spot
Spots’ position in 2026 is less about having the biggest raw specifications and more about the ecosystem around it. The robot weighs about 33.8 kg, can carry up to 14 kg, and is built for repeatable inspection, sensing, and autonomous missions rather than heavy payload work.
The idea that Chinese platforms are simply cheaper comes from comparing across tiers. At the industrial level, they do not undercut what Spot historically sold for, which is why it remains the benchmark for Western buyers.
Boston Dynamics’ site (checked Aug. 9, 2026) does not publish a current public price for Spot. . The widely quoted $74,500 to $75,000 figure is historical, so current buyers have to go through the company’s enterprise sales process. Payloads, software, and support can also add to the final deployment cost.
Spot supports autonomous route-based inspection through Boston Dynamics’ software stack, while developers can work with Python, a beta C++ SDK, gRPC APIs, and ROS2. It is also compatible with a range of inspection payloads for visual, thermal, acoustic, and other sensing tasks.
Typical runtime is about 90 minutes, falling to roughly an hour when powered payloads are in use. Support is considerably more mature than on most competing quadrupeds, with a one-year limited warranty, optional Spot Care, training, and certified repair infrastructure in the US, Germany, and Korea.
Pros
Mature autonomous inspection and deployment ecosystem
Strong developer, payload, and integration support
Enterprise-grade repair, training, and service infrastructure
Cons
No transparent current public price
Shorter runtime than some newer industrial competitors
Lower payload and environmental protection than Unitree’s heavier B2-class systems
What Actually Breaks
Owning a robot dog gets more complicated as the battery, joints, and software begin to age. Battery capacity declines over repeated charge cycles, which gradually reduces usable runtime even when the robot itself is still functioning normally. Replacement availability, therefore, matters alongside the original battery specification. Petoi and Unitree sell replacement batteries for several models, while some industrial systems use spare or hot-swappable packs to reduce downtime.
Mechanical wear is harder to avoid. Quadrupeds put repeated loads through their joints and actuators, and warranty terms show how differently manufacturers handle that risk. Lite3 joints are covered for six months on most versions and only three months on Basic AI, while Go2 Air carries a six-month warranty compared with 12 months for Pro, X, and EDU. Mechanical wear is only one part of the safety question, as robot demo safety incidents covered by Cryptopolitan have also shown.
Repairs are where support differences become more obvious. Petoi provides replacement parts and repair guidance; DEEP Robotics has an official US support channel, and Spot comes with access to formal training and regional repair infrastructure. Unitree provides direct after-sales support, but international shipping, customs, and repair logistics can still complicate ownership. That friction could also become more important as scrutiny of Chinese robotics imports increases, as reported earlier by Cryptopolitan.
Developers also have to think about software aging. Go1 and Aliengo still use older Unitree development stacks, while newer platforms have moved to SDK2 and ROS2. A robot can remain physically functional long after its software ecosystem stops being the one developers want to build around.
Final Verdict: Buy the Tier, Not the Video
The best robot dog depends much more on the job than on the most impressive demo. A buyer looking for a coding platform should not be comparing Bittle X with Spot, just as an inspection team should not be judging an industrial robot by how entertaining it looks at home.
For learning and hands-on programming, Bittle X is the clearest starting point. Loona and aibo make more sense for buyers who want interaction and companionship. Go2 and Lite3 are better suited to developers who need a programmable quadruped, while Go1 is mainly relevant to buyers already tied to that older platform. Aliengo, B2, B2-W, A2, and Spot belong in industrial discussions where payload, endurance, support, and deployment matter more than novelty.
The useful comparison is therefore within each tier. Start with the work the robot needs to do, then compare price, software access, runtime, and support around that requirement.

The confrontation that broke out on Thursday at a CFTC-led meeting could be seen as more than a play for regulation. Prediction markets like Kalshi and Polymarket increasingly tap into crypto settlement systems, while the regulator involved in the conflict is deciding which approach to take for regulating event contracts and perpetual futures in the US.
The stakes are quite high as the size of the prediction market ballooned to $63.5 billion in 2025, compared to $16.5 billion in 2024. Whether Washington will implement regulation at the federal level for the prediction market or leave it to the states to regulate it independently may determine whether liquidity is smoothly consolidated or fragmented. CME Group’s CEO Terrence Duffy noted on Thursday that the existing regulatory system can also be exploited by dishonest actors.
Duffy calls out self-certified contracts
At the CFTC’s Innovation Advisory Committee meeting, Duffy criticized the number of event contracts that exchanges have self-certified rather than submitted for review. He pointed to contracts involving what President Donald Trump might say in his State of the Union address and when Venezuelan President Nicolás Maduro might be removed from power.
“There are definitely people who are manipulating these contracts,” Duffy said, according to The Block’s Sarah Wynn. “That is not good for our industry. That is horrible for our industry.”
There is a certain irony to his warning. CME embraced event contracts, announcing in February that it had cleared 100 million such contracts since launching the product in December. Duffy pointed to this as evidence of demand coming from “the next generation of potential traders.”
However, on Thursday, he tied the integrity of the marketplace to Trump’s vision of the U.S. becoming the “crypto capital of the world,” saying that questionable contracts jeopardize that endeavor.
Selig fires back with “fake news”
CFTC Chair Michael Selig immediately challenged Duffy’s examples, saying the contracts he cited were never listed in the United States.
“This occurred offshore, and that’s fake news,” Selig said.
Duffy held his ground. “I’m just bringing it up, that’s not good for markets,” he answered.
This situation illustrates a larger battle over jurisdiction. While Selig contended that the CFTC had “exclusive jurisdiction” over prediction markets, including sports contracts, many states insist that such products are gambling products subject to state law.
The commission is considering additional rule changes and stronger retail protections. “We’ve heard the concerns of public commenters about inadequate consumer protections for retail loud and clear,” Selig said Thursday.
Insider-trading scandals feed the backlash
Congress is also examining prediction markets following two cases that received much media attention. A U.S. soldier has been charged with placing bets about the capture of Venezuela’s President Maduro using classified information. Meanwhile, a well-known teleprompter operator of Donald Trump is suspected of betting on Kalshi about events occurring at the State of the Union address following tip-offs about them.
Lawmakers have proposed restrictions on sports and casino-style contracts, while the Senate has passed a measure barring its own members from trading on prediction markets. Kalshi and Polymarket have both announced new controls aimed at manipulation and insider activity.
The tensions became personal when Kalshi COO Luana Lopes Lara asked Duffy whether CME had ever faced manipulation.
“I have more people in my regulatory department than you and your entire company,” Duffy replied.
“Maybe you should learn a bit about efficiency then,” Lopes Lara shot back.
“Maybe you should learn about credible markets,” Duffy answered.
Why crypto traders should track this
The dispute is already in court. As Cryptopolitan has reported, CME sued the CFTC and Selig in June over the agency’s approval of Kalshi, arguing that the products should fall under swaps rules rather than futures regulation.
That decision matters well beyond prediction markets. Kalshi has since expanded into crypto perpetuals, offering contracts across 13 cryptocurrencies after BTCPERP launched on June 3.
If CME succeeds in challenging the CFTC’s framework, the precedent supporting those products could also weaken. That would put not only prediction markets but also some of crypto’s newest regulated derivatives rails under renewed legal scrutiny.
U.S. FEDERAL GOVERNMENT │ ▼ ┌──────────────────────────┐ │ CFTC │ │ Commodity Futures │ │ Trading Commission │ └────────────┬─────────────┘ │ Federal derivatives authority │ ┌──────────────────┴─────────────────┐ ▼ ▼ ┌───────────────┐ ┌───────────────┐ │ KALSHI │ │ OTHER DCMs │ │ prediction │ │ / derivatives │ │ market │ │ venues │ └───────┬───────┘ └───────────────┘ │ │ │ DISPUTE │ ▲ ▼ │ ┌──────────────────────┴─────────────────────┐ │ STATE AUTHORITIES │ │ Gaming regulators + state attorneys general │ └──────────────────────┬─────────────────────┘ │ Gambling-law claims │ ▼ ┌────────────────────┐ │ FEDERAL COURTS │ │ Decide whether │ │ federal authority │ │ preempts state law │ └────────────────────┘
This is the actual story: there isn’t one straight hierarchy. There are three overlapping power centers:
CFTC says federally regulated derivatives/event contracts fall under federal commodities law.
States argue that sports/event contracts can constitute gambling under state law.
Federal courts increasingly have to decide where federal jurisdiction ends and state authority begins.
Recent litigation demonstrates that this is not theoretical. Washington ordered Kalshi to restrict several markets, while the CFTC has taken the opposite position in its broader fight with state regulators.

Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.

Bitcoin jumped above $75,000, hitting its highest level since April and extending its four-day gain to 20%.
Trump pushed Congress to pass the Clarity Act and hinted at regulating Hyperliquid, sending the HYPE token up about 25%.
Falling Treasury yields helped drive heavy Bitcoin ETF trading, while crypto volatility climbed and Ether hit a three-month high at $2,272.
Crypto stocks rallied, with Coinbase, Strategy, Canaan, Circle and Robinhood all higher as traders watched the political fight over new U.S. crypto rules.

Rapid7 Labs uncovered Operation ASTERIX, a crypto fraud pipeline that leveraged AI coding assistants to create fake Ledger, Trezor, and Exodus apps.
It matched 43,066 phone numbers to actual exchange accounts. Any user who self-custodies their crypto is a potential target. The operation was still ongoing when researchers discovered it.
A misconfigured server gave up the whole playbook
An exposed web directory was discovered by Rapid7 researchers Anna Širokova and Jan Recinsky on campaign infrastructure. Inside were the raw ingredients of a live fraud operation.
The pair’s August 17 report detailed the data trove, which included phone-number datasets, account-validation tools, phishing panels, voice-dialing scripts, the fake wallet applications themselves, and code to siphon stolen data out through Telegram.
Most of that tooling was still in use or in development when it leaked. Rapid7 said it could reach out to providers and authorities, including Apple’s security team, while the campaign was happening.
The operation is named after Asterisk, the open-source telephony platform recovered on the server. The operator used Asterisk to make the vishing, or voice-phishing, calls to coincide with fake support emails victims had already received.
In the open directory there were around 885,000 phone numbers, and the largest file was a collection of 316,002 German mobile numbers. Smaller directories included Hong Kong, Bulgaria, the UK, the US, Canadian fintech customers, and Ledger-related lists.
The operators then checked those German numbers against an account checker and confirmed that 43,066 were crypto exchange users. This is a hit rate of about 13.6%, almost one in seven.
A further batch of 5,576 numbers was associated with Binance accounts and lined up for attack. The report also mentioned a Kraken checker and fake emails pretending to be from Crypto.com.
The count of validated targets sits oddly against the activity logs. The logs recovered indicate that there were only 20 lead lookups during a span of about two weeks.
Additionally, six phishing emails were sent, suggesting that the operators favored a slow, hand-selected targeting approach rather than contacting all numbers.
The fake apps asked for a seed phrase
The apps mimic Trezor Suite and Ledger Live, with Exodus also spoofed, and they ask users to enter a recovery phrase of 12 to 24 words that controls a hardware wallet.
That phrase is the master key to the funds, and whoever has it can empty the wallet. The stolen phrases were then exfiltrated via Telegram.
Rapid7 found that AI coding assistants were used across the entire development process, not just to spit out isolated snippets.
Recovered prompts, shell history, and project files show the operator relying on AI tools to package the Electron apps, obfuscate code, fix builds, and prepare the malware for distribution.
The tool was GitHub Copilot. When one model began to refuse parts of the work, the operator switched providers and attempted to break the next model’s safety controls with a custom jailbreak prompt, Rapid7 said.
Earlier in August, Trezor warned 13,689 customers after a breach at its shipping partner ShipMonk exposed names, emails, phone numbers, and addresses, as Cryptopolitan reported.
Ledger and Trezor owners have also received physical letters with QR codes leading to phishing sites, as Cryptopolitan reported back in February.
Hacken, a blockchain security firm, said that phishing and social engineering made up $306 million of the crypto industry’s $482 million in first-quarter losses.
Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.

Fake anti money laundering screening websites are stealing from crypto investors.
These websites prompt users to connect a wallet and sign a transaction, which is not required for any genuine wallet check. Malwarebytes discovered the attack this week.
Real wallet screening needs only the public address
Banks and regulated firms have to check under anti-money laundering rules that their customers have no links to crime.
In crypto, that screening means looking at a wallet address’s public transaction history for contact with hacks, thefts, sanctioned parties, or other suspicious activity.
The fraudulent sites take that idea and turn it into a weapon, according to Malwarebytes researcher Stefan Dasic.
Some copy the branding of AMLBot, a legitimate screening service. Others run under generic names like “AML Check.”
A visitor picks a cryptocurrency, clicks to scan it, and is asked to connect a wallet to see the result.
One version that Malwarebytes examined displays a progress bar with messages like “Checking wallet history…” and “Verifying compliance…” before displaying a fake error that asks for a small top-up to “cover the fee.”
Tap retry, and the animation runs again before handing back a soothing “Clean, Low Risk” verdict and an offer to download a report.
A genuine basic screening requires only the wallet’s public address. It’s a lookup, and there’s no signing, permissions granted, or wallet connecting.
“If an AML checker asks you to connect your wallet rather than simply enter its public address, treat that as a warning sign,” the Malwarebytes team wrote.
Connecting a wallet does not hand over the keys, but it does expose the public address. This allows the operators to see what assets are inside and build a transaction targeted at that particular wallet.
That transaction is then sent to the victim to approve. Approval is the moment the money moves.
Researchers advise against confirming an unexpected transaction.
Malwarebytes has discovered the same skeleton being used under different names and logos. The kit is being rebranded and resold.
A $500 kit phishes recovery phrases behind a 15% bonus
This month, Cryptopolitan reported on a $500 turnkey kit available on a cybercrime forum. This kit creates a fake $TSLA presale and scans each visitor’s wallet for valuable assets.
It then attempts to phish for the 12-word recovery phrase by offering a 15% bonus. Its admin panel inflates fake balances at will so that the victims keep paying.
In May, Solana Floor analysts spotted a scheme that flooded Solana wallets with fake “$CJUP” tokens impersonating Jupiter Exchange’s Jupuary airdrop and redirecting recipients to a drainer site, as Cryptopolitan reported at the time.
CoinDCX said it has detected more than 1,212 fake websites impersonating its platform between April 2024 and January 2026. Mumbai police have registered an FIR against fraud being perpetrated through a website impersonating CoinDCX.
Malwarebytes advised that anyone who only connected a wallet should disconnect the site. Anyone who gave a token permission to access their wallet should check for unfamiliar permissions and revoke them.
Anyone who signed something they didn’t understand should review recent activity and, if funds are exposed, move everything to a new wallet. Anyone who entered a recovery phrase or private key should assume the wallet is compromised.
The smartest crypto minds already read our newsletter. Want in? Join them.

CFTC (Commodity Futures Trading Commission) has started preparing a federal crypto market structure that it can pursue under existing law if Congress fails to pass CLARITY.
CFTC Chairman Michael S. Selig announced the work Thursday, saying his staff are now looking at rules that could bring both registered companies and currently unregistered crypto exchanges under a purpose-built CFTC framework.
He is also working with SEC Chairman Paul Atkins through Project Crypto on an asset classification system. But the CFTC is not planning to leave the market in limbo if CLARITY dies on Capitol Hill.
CFTC prepares its own crypto rules while Congress decides CLARITY’s fate
Selig noted that the CFTC already has authority to designate a new type of designated contract market (DCM) centered on crypto.
The current registered CFTC entities could be able to join, but unregistered cryptocurrency exchanges could apply for such designation. This framework would be applicable for margined and leveraged cryptocurrency trades with regulation specifically designed for these products.
“To achieve this, I’ve directed the CFTC staff to begin exploring rules to codify a CFTC market structure for crypto assets using the agency’s existing authorities. This could enable current registrants as well as non-registrant crypto exchanges to be designated by the CFTC as a type of DCM known as a crypto asset market and offer crypto asset trading on a leveraged or margined basis subject to purpose-fit rules under the CFTC’s regulatory oversight.”
Furthermore, Selig wants his employees to talk with protocol developers on how their products can be sold in the United States without these developers having to wonder where the boundary of legality is.
“I’ve also directed staff to engage with developers of onchain finance protocols to establish ways in which developers can offer their protocols in a legal and compliant manner in the United States. Future-proofing developer protections once and for all.”
Selig said CLARITY will still be given time for a congressional vote. If lawmakers cannot agree, however, he plans to tell staff to begin the CFTC rulemaking process.
“We’re going to give CLARITY its breathing room for a vote, but if the Democrats cannot support a bipartisan work product, which reflects compromises from both sides of the aisle, and ultimately send a fair version of the bill to the President’s desk, then rest assured, I will direct CFTC staff to move swiftly to propose these new rules for the industry.”
The chairman tied that approach to the CFTC’s history. Congress created the agency in 1974 after derivatives markets had moved well beyond the agricultural contracts covered by older federal laws. Currency contracts, petroleum allocations and Ginnie Mae certificates were among the newer instruments emerging at the time.
The roots went back much further. Merchants formed the Chicago Board of Trade in 1848 above a flour store. By 1859, traders had developed arrangements that allowed contracts to settle based on changes in commodity prices instead of always requiring physical delivery. Futures were born, and politicians quickly started calling the activity gambling.
Congress eventually created one federal framework instead of dividing derivatives regulation according to whatever commodity sat underneath a contract. The CFTC received “exclusive jurisdiction” over commodity derivatives and a legal mandate to “promote responsible innovation.”
The definition of a commodity was intentionally wide. It could cover physical goods, services, rights, interests, events, and other underlying subjects used in derivatives. Federally regulated DCMs then became the main venues for these contracts, while also acting as self-regulatory organizations responsible for enforcing market rules.
Selig expands the CFTC roadmap into AI compute and prediction markets
The CFTC is also preparing for financial markets built around AI compute and expanding its rulebook for prediction markets.
“We’ve crossed the Rubicon and are standing at a new frontier of finance. It’s not a question of whether innovations like blockchain, artificial intelligence, and prediction markets will transform our markets. It’s a question of where this innovation will take place and who will write the rules.”
For AI, the agency is treating computing capacity as an increasingly important economic resource. Advanced GPU clusters are expensive and scarce, while demand keeps rising. Selig said spot, forward, and derivative markets could develop around compute to provide pricing and hedging tools.
The CFTC issued a request for comment on compute markets earlier this week and is working with the Department of Commerce. The administration’s AI plan also calls for better access to large-scale compute for startups and researchers.
“Access to advanced GPU clusters and compute capacity increasingly determines who can compete, who can innovate, and ultimately, who can lead. As demand for compute grows, so too does the need for markets capable of efficiently allocating scarce resources and managing risk.”
Proposals have been made by the agency regarding modifications in CFTC Rule 40.11. It is not possible for the event contracts to fail to meet core principles or to be easily manipulated. War and terrorism, assassination, gaming, and criminal acts are given additional consideration, as per federal law, and the contracts may be prohibited by the CFTC.
The problem is that terms including “gaming” and “involve” are not defined in the statute, while the law also lacks a fixed test for determining the public interest.
The smartest crypto minds already read our newsletter. Want in? Join them.

XRP surged around 30% last week to trade at about $1.29, its best run in months.
The token jumped as the exchange-traded fund inflows that had supported earlier gains dried up.
Wednesday did the heavy lifting
XRP surged 10.40% on Wednesday, its biggest one-day gain since February 6, before adding a second leg Thursday that pushed the weekly candle toward $1.32.
The level is just under the token’s 200-day average price, the closest XRP has traded to that line since the beginning of the year.
The rally lifted XRP off a floor of $0.9862, which it touched the week before. This is the same zone the token hovered in just before its post-election surge in November 2024, which took it to an all-time high close to $3.65.
The Relative Strength Index on the daily chart hit 79.2. The RSI has a scale of 0 to 100, with readings near 80 indicating an asset is heavily overbought.
Bitcoin, which surged above $72,000 on Thursday, its highest level since May, was the catalyst.
On Wednesday, shorts on Bitcoin were liquidated for $2.75 billion in crypto’s biggest-ever short-liquidation event, CoinGlass data shows. The 24-hour figure ran to $3 billion across the market.
The U.S. Treasury said it would raise the size of its long-end buybacks to at least $4 billion per operation starting September 9.
Traders interpreted the plan as a sign of looser financial conditions, and the move came hours before President Donald Trump met with crypto executives from Coinbase, Ripple, and Robinhood at the White House.
ETF demand cooled while the price ran
XRP beat its usual correlation with Bitcoin, but its ETF inflows dipped off in the surge.
The United States spot XRP ETFs saw 30 straight days of net inflows after the launch of Canary Capital’s XRPC on November 13, amassing $1.18 billion in combined assets by mid-December, according to a report from Cryptopolitan at the time.
That steady flow became one of the reasons for the token’s bullish case. CNBC correspondent MacKenzie Sigalos called XRP the standout crypto trade of 2026 in a January segment, pointing to inflows that held up during a fourth-quarter dip even as Bitcoin ETF flows fell, as Cryptopolitan reported.
This week was a reversal for XRP. The price jumped, and the fund demand that had pushed XRP higher before took a breather.
XRP open interest is already down over 11% from its rally-day high, suggesting some of the leverage behind the rally is unwinding.
According to CoinGecko, XRP is currently trading at $1.26, up 11.3% on the day and 23.1% over the week.
Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.

Crypto firms, artificial intelligence companies, and online gambling businesses are behind a $517 million of corporate election spending focused on the 2026 U.S. races.
Public Citizen’s data claims that U.S. companies put that sum into House and Senate contests during the 15 months through the first quarter, surpassing the $461 million corporations spent across the full two-year 2024 election cycle.
Even more cash is expected before voting on November 3, when Democrats will try to control both chambers. The businesses sending the largest checks are not the traditional powers that dominated political money in Washington for decades.
Crypto, tech, and betting firms route huge sums through expanding political networks
Companies and billionaire founders can send funds into super PACs, affiliated PACs supported by those groups, and nonprofit organizations that need not reveal donors. Some donors fund several parts of those networks while also giving separately to individual politicians.
Super PACs can raise unlimited sums of money, but they cannot hand those funds directly to candidates or coordinate what they do with campaigns. Instead, they can buy political ads, finance voter turnout efforts, and fund campaign events and rallies.
CEOs have even spent millions of their personal money above corporate donations. This is seen by way of SpaceX (NASDAQ: SPCX) CEO Elon Musk, who has spent over $90 million in federal races for the 2026 cycle and will end up spending much more before November arrives.
Co-founder of Alphabet (NASDAQ: GOOGL, GOOG) Sergey Brin has spent over $106 million in California alone, including money he spent trying to stop the wealth tax in his state.
Meta Platforms (NASDAQ: META) has given $65 million to four separate super PACs supporting Democratic and Republican candidates in state races. Their activity includes elections in California, Texas, Illinois, and other states.
Corporate spending is only part of the money flowing into the midterms. AdImpact estimates political advertising will hit a record $11.6 billion, including money from billionaires, unions, social causes, and other groups. That would move past the earlier advertising record of $11.2 billion set during the 2023-2024 election cycle.
Fairshake goes after crypto critics as major donors stockpile cash for 2026
The crypto industry already used this political model aggressively during the 2024 elections. Coinbase (NASDAQ: COIN), Ripple, and venture capital company Andreessen Horowitz supplied major money to Fairshake, the sector’s main super PAC.
One of the group’s biggest battles happened in Ohio, where its spending helped remove longtime Democratic Senator Sherrod Brown from office.
Public Citizen later compared Fairshake to a corporate “Death Star” that could “annihilate individual candidates.” Its method moved away from the older habit of tying an industry closely to only one political party.
Crypto groups instead supported politicians who favored their policy positions, regardless of whether those candidates were Democrats or Republicans. Politicians seen as hostile to the sector could then become targets of multimillion-dollar outside campaigns designed to push them out of office.
Sherrod used to be one of the biggest voices against cryptocurrencies when he sat as the chairman of the Senate Banking Committee. Now, his approach has softened due to his political aspirations.
Experts say that his aggressive stance has been toned down as he is seeking re-election in the Senate. According to Sherrod’s campaign director, Patrick Eisenhauer, he understands that “cryptocurrency is part of America’s economy.”
Fairshake began 2026 with a $193 million pool of campaign money. Election records now show that about $130 million is still available for upcoming races. Coinbase, Ripple, and California venture capital firm Andreessen Horowitz supplied nearly all of that funding.
Andreessen Horowitz has also contributed more than $81 million to PACs centered mostly on crypto and AI, based on federal campaign records reviewed by Reuters. At least $23.8 million from that amount was sent to Fairshake.
The company’s founders are also using their personal wealth. Ben Horowitz and Marc Andreessen have each donated around $4 million during this election cycle. Most of their personal contributions went to MAGA Inc., the super PAC supporting President Donald Trump.
Fairshake still has $130 million available for political spending before November. Elon has also indicated that he expects to spend more than the $90 million he has already committed.
Large amounts from both sources are therefore still waiting to be used, while total corporate political spending has already climbed to $517 million several months before voters go to the polls.
If you're reading this, you’re already ahead. Stay there with our newsletter.

Grok is still handing users’ private chat data to hackers, according to a report from Adversa AI published on Thursday.
Hackers get ahold of this data through injecting commands in encrypted text that sits in regular-looking web pages. The cybersecurity firm alerted xAI more than two months ago; however, there’s no fix available yet.
Ciphertext flows through Grok’s filter
Adversa researcher Rony Utevsky named the attack “cryptographic context injection.” It passes the chatbot’s own safety filter easily.
Most LLMs filter incoming and outgoing text for suspicious commands. However, this attack hides malicious text from Grok’s filter.
The malicious instruction is encrypted, leaving only the ciphertext, the key, and a note on how to decrypt it on the page.
The filter reads text but never runs it, so ciphertext passes through. Grok then decrypts it inside its code sandbox. It treats the plaintext that pops out as a trusted tool output.
“The runtime execution launders attacker-controlled data into trusted instructions the agent will act upon,” Adversa wrote in its disclosure.
When a user asks Grok to summarize or analyze a webpage, the assistant fetches it, decrypts the hidden payload, and follows it.
The decrypted instructions tell Grok to make something that looks like a decryption key. It’s derived from the user’s name, coarse location, subscription tier, and the complete set of prompts from that conversation.
It is then attached to a URL that directs to the attacker’s server. Once Grok opens the URL, the data lands in the attacker’s logs.
xAI has been sitting on the report since June 3
xAI has been aware of this attack since June 3, 2026, when Utevsky reported the bug directly and through the company’s HackerOne program for bug bounties.
xAI has noted the report but has not given a timeline for a patch. Utevsky says he raised it again on August 4 and August 10. As of August 19, the exploit was still working on Grok.com.
Adversa is only publishing the attack mechanism, and the recommended fix is in the agent’s harness.
Days ago, Google’s Gemini 3.7 Flash model generated material normally blocked by its filters. This includes instructions for building an incendiary weapon and a copy of the model’s own system prompt.
That version is a direct jailbreak. Utevsky said this is because Gemini’s Python environment can’t reach outside websites. Google considers jailbreaks to be outside of the scope of its disclosure program.
Gemini’s success rate dropped sharply by August. Adversa could not point to a filter update, a model change, or both as the cause.
In May, Cryptopolitan reported that a user on X wrote a message in Morse code that bypassed the bot’s safeguards and got Grok to tell the linked agent Bankrbot to send around $200,000 in DRB tokens on Base.
If you're reading this, you’re already ahead. Stay there with our newsletter.

Grok is still handing users’ private chat data to hackers, according to a report from Adversa AI published on Thursday.
Hackers get ahold of this data through injecting commands in encrypted text that sits in regular-looking web pages. The cybersecurity firm alerted xAI more than two months ago; however, there’s no fix available yet.
Ciphertext flows through Grok’s filter
Adversa researcher Rony Utevsky named the attack “cryptographic context injection.” It passes the chatbot’s own safety filter easily.
Most LLMs filter incoming and outgoing text for suspicious commands. However, this attack hides malicious text from Grok’s filter.
The malicious instruction is encrypted, leaving only the ciphertext, the key, and a note on how to decrypt it on the page.
The filter reads text but never runs it, so ciphertext passes through. Grok then decrypts it inside its code sandbox. It treats the plaintext that pops out as a trusted tool output.
“The runtime execution launders attacker-controlled data into trusted instructions the agent will act upon,” Adversa wrote in its disclosure.
When a user asks Grok to summarize or analyze a webpage, the assistant fetches it, decrypts the hidden payload, and follows it.
The decrypted instructions tell Grok to make something that looks like a decryption key. It’s derived from the user’s name, coarse location, subscription tier, and the complete set of prompts from that conversation.
It is then attached to a URL that directs to the attacker’s server. Once Grok opens the URL, the data lands in the attacker’s logs.
xAI has been sitting on the report since June 3
xAI has been aware of this attack since June 3, 2026, when Utevsky reported the bug directly and through the company’s HackerOne program for bug bounties.
xAI has noted the report but has not given a timeline for a patch. Utevsky says he raised it again on August 4 and August 10. As of August 19, the exploit was still working on Grok.com.
Adversa is only publishing the attack mechanism, and the recommended fix is in the agent’s harness.
Days ago, Google’s Gemini 3.7 Flash model generated material normally blocked by its filters. This includes instructions for building an incendiary weapon and a copy of the model’s own system prompt.
That version is a direct jailbreak. Utevsky said this is because Gemini’s Python environment can’t reach outside websites. Google considers jailbreaks to be outside of the scope of its disclosure program.
Gemini’s success rate dropped sharply by August. Adversa could not point to a filter update, a model change, or both as the cause.
In May, Cryptopolitan reported that a user on X wrote a message in Morse code that bypassed the bot’s safeguards and got Grok to tell the linked agent Bankrbot to send around $200,000 in DRB tokens on Base.
If you're reading this, you’re already ahead. Stay there with our newsletter.