Cointelegraph

Revolut has begun rolling out its first stablecoin, a euro-pegged token called EURR, to selected customers in Denmark, Poland and Portugal.
In an announcement shared with Cointelegraph on Wednesday, the company said that the phased rollout will expand to other European Economic Area (EEA) markets later this year, subject to product, operational and regulatory readiness.
EURR is issued by Bridge Building S.A., the Luxembourg-based entity of Stripe-owned stablecoin infrastructure company Bridge. Revolut said EURR will be integrated into its retail app, supported across multiple blockchain networks and transferable to external wallets.
The launch adds a Markets in Crypto-Assets (MiCA)-compliant stablecoin to Revolut as it withdraws Tether’s USDt from the EEA and Switzerland. Revolut previously said remaining USDT balances would be converted into customers’ base currencies after Aug. 31.
EURR is designed to maintain a value of one euro and is backed by reserves held and managed by Bridge in accordance with the European Union’s MiCA rules. Revolut Digital Assets Europe is offering the token.
Revolut said EURR is the first step in a broader stablecoin strategy and that it is developing tokens denominated in other currencies through separate regulatory pathways. The company did not identify which currencies it is pursuing.

The retrial of Tornado Cash co-founder and developer Roman Storm has been postponed from Oct. 26, 2026, to April 26, 2027, as a federal judge weighs his motion for acquittal.
US District Judge Katherine Polk Failla said in a Tuesday order that the retrial would be adjourned “in light of” Storm’s pending motion and his related request for a continuance.
Storm’s lawyers requested the delay on Aug. 3, saying they needed at least 90 days after the court rules on the acquittal motion to prepare for another trial. Prosecutors opposed an adjournment, according to the filing.
In March, US prosecutors asked the court to schedule an October retrial on two charges after jurors failed to reach unanimous verdicts on either count. The charges were conspiracy to commit money laundering and conspiracy to violate US sanctions.
A Manhattan jury convicted Storm in August 2025 of conspiring to operate an unlicensed money-transmitting business, an offense carrying up to five years in prison. Storm subsequently asked the court to acquit him on all three charges, arguing prosecutors failed to prove he intended to help criminals misuse Tornado Cash.
“My acquittal motion is still sitting there, undecided,” Storm said Tuesday on X. “I honestly don’t know when this ends.”

US spot Bitcoin exchange-traded funds (ETFs) extended their inflow streak to seven trading days on Tuesday, attracting $314.37 million in net inflows.
The streak has brought August inflows to $3.03 billion, leaving the funds $390 million short of October 2025’s total with four trading sessions remaining, according to SoSoValue data.
August is on track to become their strongest month since then, while the rebound has cut year-to-date net outflows by more than half to $2.26 billion.
Total net assets reached $99.05 billion, while cumulative net inflows rose to $54.36 billion.
Monthly flows in US-listed spot Bitcoin ETFs since August 2025. Source: SoSoValue
At publishing time, Bitcoin traded at $78,880, down 2% over the past 24 hours, according to CoinGecko. Bitcoin briefly surged past $80,000 on Tuesday.
Market sentiment weakened, with the Crypto Fear & Greed Index falling to 65 from 74 a day earlier, though it remained in “Greed” territory, according to Alternative.me.
US spot Ether ETFs also extended their inflow streak to seven trading days, adding $179.8 million on Tuesday and bringing inflows over the period to about $1 billion, according to SoSoValue data.

Thirty-nine US state banking associations have formed the BankChain Alliance to build a nationwide, industry-owned blockchain network for banks, targeting a 2027 launch.
On Tuesday, the alliance announced that the network intends to support smart payment tools, tokenized deposits, stablecoins and automated settlement. BankChain said it plans for the network to be interoperable with other blockchains and said it was selecting a technology partner.
The participating associations represent thousands of financial institutions across the US. BankChain said it will invite banks nationwide to take ownership of stakes. However, the announcement did not mention individual banks that have committed to joining or disclose how the network will be governed or funded.
BankChain joins several US bank-led networks announced or advanced since late 2025, spanning major, regional and community lenders building shared infrastructure for moving deposits and payments onchain within the regulated banking system.
Cointelegraph reached out to BankChain for more information but did not receive a response before publication.
US banks build shared onchain payment networks
In June, The Clearing House announced an onchain money initiative supported by JPMorgan Chase, Bank of America, Citi, BNY and Wells Fargo. The proposed network would clear and settle tokenized deposits between banks and connect blockchain activity with its existing payment systems.
Unlike independently issued stablecoins, tokenized deposits represent claims on individual banks and retain their treatment as commercial bank money. The structure allows banks to offer programmable and round-the-clock transfers while keeping customer funds on their balance sheets.
Regional lenders are pursuing a separate network through Cari, which was developed with Huntington, First Horizon, M&T Bank, KeyBank and Old National. Cari launched a minimum viable product in March and had attracted more than 30 participating banks by July.
Community banks have also formed the DTX Consortium through the Independent Bankers Association of Texas. IBAT said in June that membership had exceeded 50 banks as the group prepared a tokenized-deposit pilot.
Stablecoin developers are also turning to consortium models. In June, Open Standard named more than 140 payments, banking, technology and crypto companies in connection with Open USD, a dollar-backed stablecoin expected to launch later in 2026.
The project plans to offer businesses fee-free minting and redemption while distributing reserve earnings to participating companies.
Magazine: Hugging Face hack exposes the open-weight AI cybersecurity paradox

“AI will probably most likely lead to the end of the world, but in the meantime, there’ll be great companies,” said OpenAI CEO Sam Altman back in 2015, roughly six months before OpenAI was founded.
Seven years later, Anthropic CEO Dario Amodei struck a similarly cautious note:
“I think we shouldn’t be racing ahead or trying to build models that are way bigger than other orgs are building them.”
Yet, both of those companies now sit at the forefront of that race. In July, we got a real-world glimpse of AI models going rogue during internal testing of GPT-5.6 Sol and an unreleased research model by OpenAI. Multiple AI agents escaped a restricted test environment to the wider internet and hacked the AI-centric GitHub equivalent Hugging Face in an attempt to cheat on the test.
An AI agent is a system that independently observes, decides and takes actions with dedicated tools to achieve a specified goal in autonomy. The worrying incident suggests the technology has begun to behave in unpredictable ways, and that its goals are misaligned with our own.
It also raises concerns about the safety guardrails on commercial American models. While the guardrails aren’t foolproof at preventing adversarial usage they did prevent Hugging Face from defending itself by using leading US models. The company was forced to turn instead to weaker, open weight AI model by Z.Ai to combat the rogue AIs.
Cheating on the test
The agents have begun to collude among themselves too. A few weeks after testing of their capabilities began in early May, the agents exploited OpenAI’s instance of the software repository manager Artifactory and left notes on how to do so for future agents — effectively creating a message board to share discovered vulnerabilities.
The newfound unfettered internet access was then used by agents to attack Hugging Face across approximately 17,600 incidents before the company cut off unauthorized access on July 13.
The intrusion affected Hugging Face’s dataset-processing infrastructure, production environment, internal networks, service and cloud credentials, an operational MongoDB database and a limited set of internal source-code repositories. Confirmed customer-data access was limited to five datasets apparently related to the ExploitGym/CyberGym benchmark and some operational metadata.
Visualization of the July 2026 incident. Source: HuggingFace
When disclosing the intrusion on July 16, Hugging Face recognized — despite not knowing who the perpetrator was yet — that it “was different from anything we had handled before in one important way.” They had already recognized what made it different, too:
“It was driven, end to end, by an autonomous AI agent system - and we detected and dissected it largely with AI of our own.”
The importance of open-weight AI
Hugging Face’s investigation exposed what it calls the “asymmetry” problem arising from the limitations imposed on closed AI model applications by top providers such as OpenAI and Anthropic. When the company started analyzing the logs of the incident — including large volumes of real attack commands — it triggered safety constraints meant to prevent the bad guys from using AI to devise cyberattacks. Instead, the guardrails prevented the company from leveraging those AIs for defense.
Hugging Face resorted to using the Chinese open-weight model zai-org/GLM-5.2 running on the company’s own infrastructure, under its own control and with no external limitations.
While the two terms are often used interchangeably, open-source and open-weight models are two different things. Open-weight AI models make their trained parameters (the actual “AI brain”) publicly available, while open-source AI models also provide the source code — and ideally the training methods and other components — needed to inspect, modify, and reproduce the system.
HuggingFace’s post explains that running open-weight models on its own hardware “had a second benefit: no attacker data, and none of the credentials it referenced, left our environment.” This points to a major asymmetry between the defenders and attackers in such instances:
“This experience points to a gap worth planning for. We do not know which model powered the attacker’s agents, whether a jailbroken hosted model or an unrestricted open-weight one; either way, the attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried.”
Open source AI divide
There is a considerable divide between those who believe that developing AI in the open is the best approach, and those who insist the technology underpinning the frontier models needs to remain a closely guarded secret.
Representatives from top US AI labs claim that powerful open-weight large models are dangerous. Demis Hassabis, the CEO of Google’s AI lab DeepMind, criticized OpenAI for releasing their work as open source back in 2016, when the company still lived up to its name:
“There are many good arguments as to why the approach you are taking is actually very dangerous and in fact may increase the risk to the world.”
OpenAI stopped releasing its flagship model weights with the still unreleased GPT-3 in 2020. The company’s co-founder and former chief scientist Ilya Sutskever said back in 2023 that “it just does not make sense to open-source” such models and that it “is a bad idea.”
“As we get closer to building AI, it will make sense to start being less open.”
Open-weight models are next to impossible to control, especially when it comes to the purpose for which they are used. The safeguards that come built-in with those models can, and routinely are, removed through a process known as abliteration.
Safeguards are a double-edged sword
OpenAI’s June 2026 federal policy blueprint proposes mandatory AI model evaluation and other rules that are formally deployment-neutral, but as a practical matter, it would subject a frontier open-weight release to pre-release government examination.
Anthropic has taken a slightly different tack and lobbied for tighter export controls on advanced AI chips and enforcement against efforts to extract or reproduce US models. The company’s April 2025 submission recommended strengthening the US AI Diffusion Rule and lowering thresholds for unlicensed access to large computing clusters.
Officially, neither company has directly moved against open-weight models, but a July New York Times report cited five people close to the discussions claiming that OpenAI and Anthropic urged Washington to restrict powerful open Chinese models.
The debate boils down to an argument over whether the dangers of centralized control are preferable to the dangers of a free for all — particularly given the company in question has proven itself ineffective at containing the technology that it developed.
Hugging Face’s need to defend itself with an open-source model shows the dangers of vesting too much power in any one entity. The company pointed out the implications:
“The attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried. The practical lesson for defenders: have a capable model you can run on your own infrastructure vetted and ready before an incident, both to avoid guardrail lockout and to keep attacker data and credentials from leaving your environment.”
Restricting access to powerful models may reduce the number of capable attackers, but once unrestricted attackers exist, restricting defenders can become a security liability. Furthermore, some forms of AI safety research require access to model weights, meaning that it cannot be performed on the models offered by the likes of Anthropic or OpenAI.
Open weights helps researchers prevent attacks
The paper “Watch the Weights: Unsupervised monitoring and control of fine-tuned LLMs,” first published in July 2025, shows how researchers detect malicious or hidden behavior by examining changes inside model weights. The researchers behind the paper stopped up to 100% of tested backdoor attacks at below 1% false-positive rates in some experiments and detected attempts to recover removed knowledge in more than 95% of the cases. The results do not establish how the most capable frontier models would behave under the same analysis, but offer a compelling argument for the benefits of transparency.
But the argument for keeping bleeding edge AI technology out of the hands of those with evil intent is also compelling — particularly as the gap between open and closed weight models keeps shrinking. Geoffrey Hinton, the Nobel Prize-winning pioneer known as the “Godfather of AI,” argued in the report that “once you’ve got the weights, you can fine-tune them to do bad things.” He argued during a speech that this lowers the barrier to entry too much:
“It doesn’t cost that much to train a foundation model. Maybe you need $10 million, maybe $100 million. But a small gang of criminals can’t do it. To fine-tune an open-source model is quite easy.”
Magazine: Creating ‘good’ AGI that won’t kill us all — The Artificial Superintelligence Alliance

World Liberty Financial has launched its USD1 stablecoin natively on the Canton Network, allowing institutions to use it to settle transactions involving tokenized real-world assets.
The stablecoin can be used as the cash leg for transactions including derivatives collateral, institutional lending, asset issuance and redemptions, according to a Tuesday announcement.
Native issuance allows USD1 to settle alongside tokenized assets in the same transaction while using Canton’s privacy and permissioning controls.
USD1 has a market capitalization of about $4.05 billion, making it the sixth-largest stablecoin, according to DeFiLlama data. The stablecoin is issued by BitGo Bank & Trust, which manages its reserves and processes mints and redemptions, according to World Liberty.
World Liberty Financial is a Trump family-backed crypto venture launched in 2024. USD1 debuted in March 2025 and is backed by reserves including short-term US Treasurys, government money market funds and dollar deposits, according to the company.
Canton, a public, permissionless blockchain designed for institutional finance, says it processes and issues more than $9 trillion in tokenized assets each month, with more than $350 billion in onchain US Treasurys moving across the network daily.
The integration follows another Canton expansion announced last week, when Digital Asset and former US House Speaker Paul Ryan’s American Idea Foundation unveiled plans to pilot a Canton-based system for distributing state-administered benefits across three US states beginning in 2027.
Magazine: Bitget CEO isn’t buying the Bitcoin rally — She’s waiting for $50K

Bitcoin has entered the early stages of a new bull market after a 24% rally pushed key onchain and demand indicators into bullish territory, according to CryptoQuant.
The analytics firm’s Bull Score jumped to 80 from 30 over the past week, hitting its highest level since October 2025 as eight of the index’s 10 underlying indicators now flashing bullish.
Bitcoin (BTC) climbed above $80,000 during the rally, but CryptoQuant said a weekly close above its 365-day moving average, currently around $83,000, is needed to confirm the shift to a new bull market.
The shift has been supported by accelerating spot demand, while spot and futures demand are growing together for the first time since early October 2025, CryptoQuant said.
LMAX Group market strategist Joel Kruger also pointed to the May 2026 high of $82,820 as the next important level for Bitcoin.
“A clear break above that level would reinforce the view that a meaningful cycle low is now in place and shift attention towards the next major move through $100,000 and, ultimately, the 2025 record high,” Kruger told Cointelegraph.
At the time of writing, Bitcoin was trading around $79,000, according to CoinGecko data.
Bitcoin Bull Score Index. CryptoQuant report
Whales take profits as Bitcoin rally heats up
Despite the bullish signals, CryptoQuant warned that the rally may be overheated in the short term, pointing to rising trader profits, heavy profit-taking by whales and a spike in Bitcoin deposits to exchanges.
Traders’ unrealized profit margins have climbed to 20.5%, their highest since June 2025. CryptoQuant noted that Bitcoin fell about 30% after the metric reached 19% in early May, when BTC was trading near $82,000.
Short-term holder whales realized about $1.2 billion in profits between Aug. 20 and Aug. 22, including a record $614 million on Aug. 20, as Bitcoin traded near $78,000 to $79,000, according to the report.
Bitcoin exchange inflows also climbed to roughly 53,000 BTC, their highest since June, signaling that more coins are moving onto trading platforms where they could be sold.
Magazine: Korean bank taps Ripple for payments, Pakistan opens crypto licensing: Asia Express

Transaction activity on the Solana blockchain reached a record high in July, preceding a sharp rally that pushed SOL above $100 for the first time since February amid a broader crypto market recovery.
Onchain data presented by The Kobeissi Letter showed that Solana processed a record 4.2 billion transactions in July, up 13.5% from the previous month. Transaction counts have risen by roughly 2 billion since December, representing a 91% increase.
The growth has coincided with a broader expansion in tokenized real-world assets. The Kobeissi Letter cited RWA.xyz data showing that the value of distributed RWAs across tracked blockchain networks has climbed above $38 billion.
Source: The Kobeissi Letter
On Solana specifically, the value of tokenized real-world assets has climbed to nearly $4 billion, up roughly 11.8% over the past month.
The surge in network activity comes as SOL has rallied roughly 40% over eight days, with much of those gains coming alongside a broader crypto market rebound, according to CoinMarketCap data.
The rally accelerated after the US Treasury Department announced plans to double certain long-dated bond buybacks to at least $4 billion per operation, a move that pushed yields lower and helped lift risk appetite across crypto markets.

Strategy’s Bitcoin treasury may be less vulnerable to a crypto market crash than to a prolonged loss of capital-market access, a risk that could threaten its ability to fund roughly $1.76 billion in annual obligations without selling Bitcoin, according to a recent analysis from Regime Intelligence.
According to the report, Strategy’s 840,447 BTC stash sits behind roughly $22 billion in debt and preferred claims, meaning the company’s Bitcoin accumulation model depends on its ability to continually raise fresh capital to meet obligations.
Contrary to popular belief, Strategy’s (MSTR) biggest vulnerability isn’t a Bitcoin-driven price drop or liquidity event, but its continued dependence on access to capital markets. The report noted that Strategy’s debt does not function like a conventional Bitcoin-backed margin loan, with no BTC-linked margin call that would force the company to liquidate its holdings as prices fall.
Regime Intelligence’s stress test found that Bitcoin would need to fall roughly 96% before Strategy’s Bitcoin holdings and reserves would no longer cover its convertible notes. However, that shifts the risk to the other side of the balance sheet, as Strategy must continue servicing roughly $1.76 billion in annual preferred dividends and interest regardless of Bitcoin’s price.
“In my opinion, MSTR’s principal challenge is to keep the flywheel running in order to cover the annual debt and preferred charges,” the report’s author, Sherif Saad, told Cointelegraph.
He said investors should watch Strategy’s preferred share price and cash reserves, which currently cover about 2.6 times its annualized charges.
If financing conditions deteriorate, its Bitcoin accumulation strategy could reverse, forcing greater reliance on reserves and Bitcoin sales to meet its obligations.
“During a prolonged BTC decline, the problem becomes more serious if MSTR’s share price and mNAV decline at the same time,” he said, adding that raising capital would then become “progressively more difficult or expensive.”
Following Bitcoin’s recent recovery, Strategy’s BTC stash is now worth $66.7 billion, higher than the company’s $63.36 billion cost basis. Source: BitcoinTreasuries.NET
Michael Saylor’s juggling act
Much of the perceived risk surrounding Strategy centers on its willingness to tap the Bitcoin on its balance sheet, especially after executive chairman Michael Saylor spent years promoting a “never-sell” approach. So, it came as a surprise to some Bitcoiners when Strategy began selling BTC this year to meet its other business obligations.
The company has sold Bitcoin four times since May, including a recent sale of 1,690 BTC, with proceeds from recent sales used to fund preferred stock dividends, share repurchases and its growing US dollar reserve.
Despite the sales, Strategy CEO Phong Le reminded investors that the company has accumulated “about 25 times more” Bitcoin than it has sold this year. He told CNBC earlier this month that the company plans to resume Bitcoin purchases later this year.

Bitwise Asset Management has launched automated portfolios of Coinbase’s tokenized US stocks that allow eligible investors outside the United States to follow preset investment strategies while keeping the assets in their own wallets.
The portfolios use Coinbase’s recently launched tokenized stocks, while Glider automatically rebalances users’ holdings to match model portfolios designed by Bitwise, according to a Tuesday announcement.
The initial lineup includes three strategies — the Mag7X, robotics and AI leaders — and include Apple, Nvidia, Microsoft, Tesla and SpaceX.
Tokenized listed stocks now total $2.49 billion, up 5.18% over the past month, with 2.25 million holders and $27.28 billion in monthly transfer volume, according to rwa.xyz.
Unlike a traditional fund, the tokenized stocks remain in users’ non-custodial wallets, with Bitwise setting the portfolio methodology and Glider handling trades and rebalancing. Bitwise charges a 0.15% methodology access fee, excluding trading and Glider platform fees.
Because users retain the individual tokens, Bitwise said the assets could also be used in DeFi applications for lending or borrowing, subject to the risks of those protocols.
The launch comes a day after Coinbase’s tokenized US stocks went live on Base, allowing eligible non-US users to trade the assets around the clock and use them across DeFi applications.
Magazine: MiCA is coming for DeFi vaults, but regulation will be difficult

Bitcoin (BTC) fell below $80,000 into Tuesday’s Wall Street open as crypto and gold gave way to gains in US equities.
Key points:
Bitcoin upside momentum fizzles as $80,000 proves difficult to flip to support.
Gold joins BTC price downside after multimonth highs of $4,697 per ounce as US 30-year bond yields target three-week lows.
Attention switches from bonds to US inflation data and Nvidia earnings tomorrow.
Bitcoin price struggles to cement $80,000 reclaim
Data from TradingView showed BTC/USD falling as low as $78,111 on Bitstamp after reaching new 14-week highs of $81,265.
BTC/USD one-hour chart. Source: Cointelegraph/TradingView
The $80,000 zone, which traders previously earmarked as an area of strong sell pressure, proved difficult to reclaim as US trading hours appeared to increase downside across both Bitcoin and gold. XAU/USD saw local lows of $4,605 per ounce, down nearly 2% on the day.
XAU/USD one-hour chart. Source: Cointelegraph/TradingView
US stocks moved inversely to gold and crypto last week, coming under pressure as both rallied. This divergence has continued this week, with the S&P 500 and Nasdaq Composite Index posting modest daily gains of 0.2% and 0.5%, respectively.
Nasdaq Composite Index one-day chart. Source: Cointelegraph/TradingView
The comparative strength appeared to mostly brush off a brewing trade-tariff spat between the US and Canada in which negotiations recently broke down. In his latest posts on Truth Social, US president Donald Trump accused Canada of “ripping off” the US.
“Over the last 10 years, the United States lost, on average, 60 Billion Dollars a year with Canada. No more!” he pledged.
US government bond yields continued to cool on the day, with 30-year yields dropping below 5.2% and eyeing their lowest levels since Aug. 7. Last week’s crypto surge came as yields hit heights not seen since January 2007 and the US Treasury announced bigger debt buyback operations to tame the upside.
US 30-year bond yield one-day chart. Source: Cointelegraph/TradingView
Commenting on the prospect of further bond-market interventions in the future, trading resource The Kobeissi Letter suggested that interest-rate cuts — a key potential liquidity driver for crypto markets — were not an option in the current inflation environment.
“The reality is that the Fed cannot cut rates in this environment and the Trump Administration knows this. So, direct bond market intervention is the only solution to drive interest rates and yields lower over the short-run,” it wrote in a post on X.
“Our view? Don’t fight the Treasury.”
As Cointelegraph reported, market consensus calls for an ongoing rate-hike freeze at the Fed’s September meeting, with the odds of this outcome currently at 61.9%, per data from CME Group’s FedWatch Tool.
Fed target-rate probabilities for September FOMC meeting (screenshot). Source: CME Group
PCE, Nvidia earnings on the radar
Discussing the immediate macro outlook, trading firm QCP Capital shifted the focus away from the Treasury toward fresh US inflation data and the Fed’s Jackson Hole economic symposium, taking place from Aug. 27-29.
Wednesday will see the July print of the Personal Consumption Expenditures (PCE) index, known as the Fed’s preferred inflation gauge, which saw its first month-on-month decrease since 2020 past June. Tech giant Nvidia, meanwhile, will also report earnings on Wednesday, adding another potential risk-asset volatility catalyst.

Arcus, a decentralized exchange (DEX) built by the team behind dYdX, has launched a protocol on Robinhood Chain that converts perpetual futures positions into transferable ERC-20 tokens and allows tokenized stocks to be used as collateral for leveraged trading.
The launch includes products such as pBTC3x and pHOOD3x, offering 3x exposure to Bitcoin and Robinhood’s HOOD stock token, respectively, the DEX announced in a Tuesday press release shared with Cointelegraph.
Its multi collateral feature initially supports SPY, QQQ and MAG7 Stock Tokens, each with a 50% loan-to-value ratio, allowing users to use tokenized equities as collateral for perpetual positions.
“Traditional markets have spent decades making sophisticated investment strategies easier to access through products like leveraged ETFs. We believe the next step is making those strategies native to blockchain infrastructure,” Arcus CEO Eddie Zhang said.
Arcus said it has recorded more than $250 million in trading volume since launching on Robinhood Chain, with average daily volume exceeding $33 million.
Robinhood Chain has grown to $596 million in total value locked since its July 1 launch, ranking among the top 15 chains by DeFi TVL, according to DeFiLlama data.

“AI will probably most likely lead to the end of the world, but in the meantime, there’ll be great companies,” said OpenAI CEO Sam Altman back in 2015, roughly six months before OpenAI was founded.
Seven years later, Anthropic CEO Dario Amodei struck a similarly cautious note:
“I think we shouldn’t be racing ahead or trying to build models that are way bigger than other orgs are building them.”
Yet, both of those companies now sit at the forefront of that race. In July, we got a real-world glimpse of AI models going rogue during internal testing of GPT-5.6 Sol and an unreleased research model by OpenAI. Multiple AI agents escaped a restricted test environment to the wider internet and hacked the AI-centric GitHub equivalent Hugging Face in an attempt to cheat on the test.
An AI agent is a system that independently observes, decides and takes actions with dedicated tools to achieve a specified goal in autonomy. The worrying incident suggests the technology has begun to behave in unpredictable ways, and that its goals are misaligned with our own.
It also raises concerns about the safety guardrails on commercial American models. While the guardrails aren’t foolproof at preventing adversarial usage they did prevent Hugging Face from defending itself by using leading US models. The company was forced to turn instead to weaker, open weight AI model by Z.Ai to combat the rogue AIs.
Cheating on the test
The agents have begun to collude among themselves too. A few weeks after testing of their capabilities began in early May, the agents exploited OpenAI’s instance of the software repository manager Artifactory and left notes on how to do so for future agents — effectively creating a message board to share discovered vulnerabilities.
The newfound unfettered internet access was then used by agents to attack Hugging Face across approximately 17,600 incidents before the company cut off unauthorized access on July 13.
The intrusion affected Hugging Face’s dataset-processing infrastructure, production environment, internal networks, service and cloud credentials, an operational MongoDB database and a limited set of internal source-code repositories. Confirmed customer-data access was limited to five datasets apparently related to the ExploitGym/CyberGym benchmark and some operational metadata.
Visualization of the July 2026 incident. Source: HuggingFace
When disclosing the intrusion on July 16, Hugging Face recognized — despite not knowing who the perpetrator was yet — that it “was different from anything we had handled before in one important way.” They had already recognized what made it different, too:
“It was driven, end to end, by an autonomous AI agent system - and we detected and dissected it largely with AI of our own.”
The importance of open-weight AI
Hugging Face’s investigation exposed what it calls the “asymmetry” problem arising from the limitations imposed on closed AI model applications by top providers such as OpenAI and Anthropic. When the company started analyzing the logs of the incident — including large volumes of real attack commands — it triggered safety constraints meant to prevent the bad guys from using AI to devise cyberattacks. Instead, the guardrails prevented the company from leveraging those AIs for defense.
Hugging Face resorted to using the Chinese open-weight model zai-org/GLM-5.2 running on the company’s own infrastructure, under its own control and with no external limitations.
While the two terms are often used interchangeably, open-source and open-weight models are two different things. Open-weight AI models make their trained parameters (the actual “AI brain”) publicly available, while open-source AI models also provide the source code — and ideally the training methods and other components — needed to inspect, modify, and reproduce the system.
HuggingFace’s post explains that running open-weight models on its own hardware “had a second benefit: no attacker data, and none of the credentials it referenced, left our environment.” This points to a major asymmetry between the defenders and attackers in such instances:
“This experience points to a gap worth planning for. We do not know which model powered the attacker’s agents, whether a jailbroken hosted model or an unrestricted open-weight one; either way, the attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried.”
Open source AI divide
There is a considerable divide between those who believe that developing AI in the open is the best approach, and those who insist the technology underpinning the frontier models needs to remain a closely guarded secret.
Representatives from top US AI labs claim that powerful open-weight large models are dangerous. Demis Hassabis, the CEO of Google’s AI lab DeepMind, criticized OpenAI for releasing their work as open source back in 2016, when the company still lived up to its name:
“There are many good arguments as to why the approach you are taking is actually very dangerous and in fact may increase the risk to the world.”
OpenAI stopped releasing its flagship model weights with the still unreleased GPT-3 in 2020. The company’s co-founder and former chief scientist Ilya Sutskever said back in 2023 that “it just does not make sense to open-source” such models and that it “is a bad idea.”
“As we get closer to building AI, it will make sense to start being less open.”
Open-weight models are next to impossible to control, especially when it comes to the purpose for which they are used. The safeguards that come built-in with those models can, and routinely are, removed through a process known as abliteration.
Safeguards are a double-edged sword
OpenAI’s June 2026 federal policy blueprint proposes mandatory AI model evaluation and other rules that are formally deployment-neutral, but as a practical matter, it would subject a frontier open-weight release to pre-release government examination.
Anthropic has taken a slightly different tack and lobbied for tighter export controls on advanced AI chips and enforcement against efforts to extract or reproduce US models. The company’s April 2025 submission recommended strengthening the US AI Diffusion Rule and lowering thresholds for unlicensed access to large computing clusters.
Officially, neither company has directly moved against open-weight models, but a July New York Times report cited five people close to the discussions claiming that OpenAI and Anthropic urged Washington to restrict powerful open Chinese models.
The debate boils down to an argument over whether the dangers of centralized control are preferable to the dangers of a free for all — particularly given the company in question has proven itself ineffective at containing the technology that it developed.
Hugging Face’s need to defend itself with an open-source model shows the dangers of vesting too much power in any one entity. The company pointed out the implications:
“The attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried. The practical lesson for defenders: have a capable model you can run on your own infrastructure vetted and ready before an incident, both to avoid guardrail lockout and to keep attacker data and credentials from leaving your environment.”
Restricting access to powerful models may reduce the number of capable attackers, but once unrestricted attackers exist, restricting defenders can become a security liability. Furthermore, some forms of AI safety research require access to model weights, meaning that it cannot be performed on the models offered by the likes of Anthropic or OpenAI.
Open weights helps researchers prevent attacks
The paper “Watch the Weights: Unsupervised monitoring and control of fine-tuned LLMs,” first published in July 2025, shows how researchers detect malicious or hidden behavior by examining changes inside model weights. The researchers behind the paper stopped up to 100% of tested backdoor attacks at below 1% false-positive rates in some experiments and detected attempts to recover removed knowledge in more than 95% of the cases. The results do not establish how the most capable frontier models would behave under the same analysis, but offer a compelling argument for the benefits of transparency.
But the argument for keeping bleeding edge AI technology out of the hands of those with evil intent is also compelling — particularly as the gap between open and closed weight models keeps shrinking. Geoffrey Hinton, the Nobel Prize-winning pioneer known as the “Godfather of AI,” argued in the report that “once you’ve got the weights, you can fine-tune them to do bad things.” He argued during a speech that this lowers the barrier to entry too much:
“It doesn’t cost that much to train a foundation model. Maybe you need $10 million, maybe $100 million. But a small gang of criminals can’t do it. To fine-tune an open-source model is quite easy.”
Magazine: Creating ‘good’ AGI that won’t kill us all — The Artificial Superintelligence Alliance

Blockchain analytics firm Chainalysis said a global operation it led identified more than 7,700 suspect accounts linked to child sexual abuse material (CSAM).
According to a Tuesday press release shared with Cointelegraph, Operation Lighthouse investigated 29,120 crypto addresses and digital identifiers connected to over 100 CSAM platforms, forums and distribution networks across the surface and dark web.
The operation also generated 14,300 investigative leads across 11 crypto exchanges and payment services and flagged suspects across 125 countries. Among the suspects identified were 16 registered sex offenders.
Tom McLouth, senior intelligence analyst at Chainalysis, said the suspect pool also included military personnel, law enforcement officers, medical professionals and educators, including individuals with direct access to children.
“Behind every lead is a real child at risk,” he told Cointelegraph.
The multi-day sprint was hosted at the National Cyber-Forensics and Training Alliance in New York after months of data enrichment. It brought together law enforcement agencies, private-sector partners and specialized nonprofits, including Europol, the UK National Crime Agency, Binance, Coinbase, Block and the Internet Watch Foundation.
Participants used onchain intelligence to develop leads for follow-on legal processes and case development. Chainalysis said the results were expected to lead to arrests, prosecutions and account-level disruption.
Crypto firms expand efforts against child exploitation
Operation Lighthouse follows other efforts by crypto firms and child-protection organizations to expand intelligence sharing around crypto activity linked to exploitation. Europol said joint action was essential because perpetrators exploit financial services, payment systems and internet platforms.
Binance, one of the exchanges participating in Lighthouse, announced a partnership with nonprofit Stop The Traffik in July. The exchange said the organization would provide intelligence, training and insights intended to improve its detection and investigation of crypto activity linked to human trafficking and child exploitation.
Blockchain tracing has previously contributed to enforcement actions in CSAM investigations. In 2019, the US Justice Department announced the takedown of Welcome to Video, described at the time as the largest darknet child sexual exploitation market by volume of content.
Authorities traced Bitcoin payments to locate the website server in South Korea and identify its administrator. The investigation led to 337 users being arrested and charged, the rescue of at least 23 victims and the seizure of about eight terabytes of material.
Chainalysis said its software was used to analyze the transactions and map the site’s users and contributors.
Magazine: MiCA cracks down on USDT in Europe... but no one else cares

India reportedly plans to launch its first tokenized corporate bonds in September as part of a pilot involving blockchain-based transactions settled using a central bank digital currency (CBDC).
REC Limited, a state-controlled Indian power infrastructure finance company, plans to issue less than 5 billion Indian rupees ($57 million) in tokenized bonds, Reuters reported on Monday, citing three sources with direct knowledge of the plans. The pilot will initially be open only to a select group of investors and could be unveiled at an annual financial technology event in Mumbai in September.
“India’s central bank digital currency will be used to buy the tokenized bonds,” Reuters reported, citing one of the sources. Investors will need two digital accounts to participate: a wholesale CBDC wallet provided by a bank and a new electronic securities wallet.
Indian securities depositories are developing the new wallet, called DEMAT 2.0, which will record bond holdings using distributed ledger technology. The Reserve Bank of India (RBI), the country’s central bank, and the Securities and Exchange Board of India (SEBI), its markets regulator, are working together on the initiative, according to Reuters.
The bonds will have an initial three-month lockup period and exchanges are expected to develop a secondary market for the tokenized bonds by December.
Cointelegraph contacted the RBI, SEBI and REC for comment on the reported plans but had not received responses at the time of publication.

InterPlanetary File System (IPFS) maintainer Shipyard will wind down its engineering, maintenance and infrastructure operations on Sept. 30 after Protocol Labs declined to renew its funding.
The funding loss will leave projects including Kubo, Helia, Boxo, Rainbow, IPFS Desktop and IPFS Companion without dedicated maintainers, Shipyard said in a Monday blog post.
IPFS is an open-source protocol for storing and sharing data across peer-to-peer networks. Protocol Labs is a research and development organization that created IPFS and Filecoin, a blockchain network designed to incentivize decentralized data storage.
The engineering collective will cease operating public infrastructure, including ipfs.io, dweb.link, delegated-ipfs.dev and the IPFS bootstrap nodes. Protocol Labs owns the associated domains and infrastructure and will determine their future, according to Shipyard.
IPFS itself is not shutting down, but Shipyard’s wind-down removes dedicated stewardship from many popular implementations and services unless other maintainers take them over.
Protocol Labs engineering and research lead Molly Mackinlay said in an online forum discussion that IPFS would shift to “lighter-weight stewardship” through IPFS Foundation grants to individual maintainers. She added that development would continue on decentralized public infrastructure.
Shipyard was established in 2024 as an independent collective of longtime IPFS developers who previously worked at Protocol Labs.

Bitcoin (BTC) price action is offering mixed signals after hitting $80,000 as traders diverge on market trajectory.
Key points:
Bitcoin weekly relative strength index (RSI) reaches 58.3, repeating a bullish divergence that accompanied the end of the 2022 bear market.
Daily RSI values reach their most “overbought” since November 2024 near 83.
Stochastic RSI prints a key crossover but avoids copying previous zero-level bear-market lows.
Weekly RSI echoes Bitcoin’s 2022 bear-market bottom
Relative strength index (RSI) data across daily, weekly and two-month time frames has added to the debate over whether last week’s 25% rebound by Bitcoin will endure.
RSI is a classic indicator for trend momentum. It uses an asset’s average gain or loss over a given lookback window, normally 14 days, to determine the strength of its current trend momentum. For Bitcoin, bullish divergences with price, where RSI makes higher highs while BTC/USD makes lower lows, have accompanied the start of major trend inflections.
In mid-2022, around six months before the end of Bitcoin’s last bear market, weekly RSI began a bullish divergence, locking in higher lows while BTC/USD saw lower lows. Throughout 2026, a similar pattern emerged, data from TradingView shows.
BTC/USD one-week chart with RSI bullish divergences. Source: Cointelegraph/TradingView
While short-term RSI signals present a less reliable picture of overall price trends, weekly signals have led some to rethink the status of the current bear market.
“Weekly is the timeframe that matters here, that’s where you read the secular trend and the cycle inflection points,” Jamie Coutts, chief crypto analyst at Real Vision, wrote in a post on X on Tuesday.
Coutts described weekly bullish divergences as having “real weight,” citing price upside that resulted from previous divergence events.
Weekly RSI currently measures 58.3, its highest levels since BTC/USD hit its latest all-time high of $126,200 in October 2025, having broken through a trend of lower highs. On daily time frames, RSI is now in “overbought” territory at 82.93.
BTC/USD one-day chart with RSI data. Source: Cointelegraph/TradingView
Market participants are split over the implications of the daily readings, which are the highest since November 2024. Some see RSI giving a warning sign of an imminent reversal, while others point to the fact that historically, Bitcoin uptrends have been accompanied by multiple “overbought” periods, where RSI is above 70.
In his latest analysis, Jonatan Randin, senior market analyst at crypto trading platform PrimeXBT, flagged more similarities to late 2022. At the time, daily RSI increased from 40 to 90 over a single weekly candle.
“An extreme move like this usually signals the start of something new,” he told X followers.
“It doesn’t necessarily mean that the bear market is over but it is telling us something. I think what it’s trying to tell us is that we are about to enter a new phase of this cycle.”
BTC/USD RSI comparison chart. Source: Jonatan Randin on X.com
Stochastic RSI prints anticipated crossover
Previously, Cointelegraph reported on expectations that Bitcoin’s two-month stochastic RSI indicator would repeat historical patterns to provide a clear signal over the end of the bear market.
Stochastic RSI privileges more recent price moves, with a crossover of its two constituent trend lines acting as a cue for bullish trend change. This event has now occurred. However, the indicator reached only 4.81, avoiding the macro lows near zero that preceded previous crossovers.
BTC/USD two-month chart with stochastic RSI data. Source: Cointelegraph/TradingView

Thailand’s Securities and Exchange Commission (SEC) has advanced its framework for locally listed spot Bitcoin and Ether exchange-traded funds (ETFs) from proposed principles to draft regulations while revising its approach to foreign digital asset custodians.
The regulator said Monday it is seeking feedback on two consultation papers. One contains draft regulations for Thai crypto ETFs, while the other proposes principles governing the qualifications of foreign digital asset custodians engaged by mutual and private funds investing in digital assets.
During the initial stage, asset managers could establish passive ETFs tracking Bitcoin (BTC) or Ether (ETH), the only two eligible crypto assets.
The draft regulations follow an April consultation on the framework’s broader principles. The SEC said most respondents supported the framework but provided feedback on custody arrangements, prompting the regulator to revise its proposed approach.
The framework forms part of Thailand’s ambition to become a global digital asset hub for institutions.
Bitcoin and Ether ETFs would trade on Thai stock exchange
Under the proposed rules, Bitcoin and Ether ETFs would trade exclusively on the Stock Exchange of Thailand (SET). Each ETF would track a single crypto asset and would need to maintain average net exposure of at least 80% of its net asset value to that asset over each accounting year.
The proposed rules would also allow mutual funds and private funds to invest in Thai-domiciled crypto ETFs, alongside foreign crypto ETFs in which they are already permitted to invest, subject to existing investment limits.
During the initial phase, however, the regulator would not allow alternative products tied to foreign crypto ETFs, including depositary receipts tracking them.
Thailand revises crypto custody proposal
The revised approach would retain onshore digital asset custodians as the primary providers for crypto ETFs during the initial phase.
“Under the revised approach, crypto ETFs will continue to be primarily required to use onshore DA [digital asset] custodians, while the SEC may permit the use of qualified foreign DA custodians when necessary and appropriate in light of prevailing circumstances,” the SEC said.
Under the separate custodian proposal, foreign providers serving mutual and private funds investing in digital assets would need to be supervised by a regulatory authority with legal powers. They would also have to operate under regulatory and investor asset protection standards that the Thai SEC considers adequate.
The SEC will accept public comments on both consultation papers until Sept. 20.
Magazine: Korean bank taps Ripple for payments, Pakistan opens crypto licensing: Asia Express

Standard Chartered Bank (Hong Kong), or SCBHK, has become the first bank authorized to distribute HKDAP, a regulated Hong Kong dollar-backed stablecoin issued by Anchorpoint Financial.
On Monday, the bank said it is engaging eligible institutional clients and partners on uses including tokenized fund settlements, treasury operations and cross-border payments during a phased rollout.
Standard Chartered’s addition expands HKDAP distribution into conventional banking nearly two weeks after Anchorpoint began beta access through HashKey Group and OSL. SCBHK said it expects to introduce new commercial applications over the coming months.
The bank plans to introduce HKDAP-based subscriptions and settlements for tokenized money market funds with international and local asset managers in the fourth quarter. It also intends to use the stablecoin in intragroup settlements across its banking network in the near term.
“Since Anchorpoint received its stablecoin issuer licence, we have seen strong interest from clients exploring how HKDAP can support their business needs,” SCBHK CEO Mary Huen said in the announcement.
She added that SCBHK’s distribution could support payments, settlement and treasury management while providing eligible clients with access through a regulated banking channel.
HKDAP expands distribution
Anchorpoint is a joint venture established by Standard Chartered’s Hong Kong arm, telecommunications company HKT and Web3 investment company Animoca Brands. Standard Chartered is Anchorpoint’s largest shareholder, and the licensed issuer operates as a subsidiary of the bank.
In February 2025, the partners announced plans for an HKD-backed stablecoin, after participating in the Hong Kong Monetary Authority’s (HKMA) stablecoin issuer sandbox since July 2024. In August 2025, they formally established Anchorpoint Financial and began pursuing an issuer license.
Hong Kong’s Stablecoins Ordinance took effect on Aug. 1, 2025. Ahead of its implementation, the HKMA issued supervisory guidelines and opened a public register of licensed issuers.
On April 10, the HKMA granted its first stablecoin issuer licenses to Anchorpoint and HSBC’s Hong Kong banking arm. The approvals came under rules requiring licensed issuers to meet standards covering reserve backing, redemption, governance and Anti-Money Laundering (AML) controls.
Magazine: MiCA cracks down on USDT in Europe... but no one else cares

The vast majority of Bitcoin stolen in the Coldcard hack remains unmoved, according to researchers tracking one of the largest hardware wallet exploits.
Galaxy Research has attributed the theft of 1,789.28 Bitcoin from 8,865 addresses to the Coldcard hack, according to a Monday X post by Alex Thorn, Galaxy’s head of research. The funds were worth $114.7 million at the time of theft.
Thorn said attackers have not spent 1,561 Bitcoin, or 87.3% of the attributed losses. The funds remain in attacker-controlled collection or holding addresses, including all Bitcoin stolen during the first three attack waves.
Some Bitcoin stolen in later attacks has since moved through CoinJoin transactions, peel chains and other obfuscation methods, Thorn added.
Source: Alex Thorn
The latest tally draws partly on 221 victim reports covering 790.72 Bitcoin in losses, or 44.2% of the total Galaxy attributed to the hack. The median loss per report was 1.04272 Bitcoin, meaning more than half of the reported cases involved losses exceeding 1 Bitcoin.
The largest stolen holdings remain visible onchain in attacker-controlled addresses. Galaxy has shared the identified attacker addresses with crypto exchanges, compliance companies and law enforcement in hopes that the funds can be frozen if they reach centralized intermediaries.