oAdam

AI 智能体正在改变网络攻击的执行方式：它不只执行命令，还会自己试错、复盘，再换方法继续。

公开案例已经出现了令人警惕的信号。

韩国一家金融机构遭入侵，约 6.7 万人的个人信息外泄，攻击中使用了名为 ARTEX 的智能体渗透框架。

另一场供应链攻击里，同一套框架尝试登录超过 2250 个租户，部分攻击在不到三小时内就获得了特权云访问权限。泰国财政部相关攻击中，Hermes 智能体以无人值守模式运行，自动枚举内网主机、翻查文件、寻找提权机会。

Wiz 对相关攻击活动的研究，揭示了几个值得识别的行为特征：

* 高频生成脚本： 一起攻击中，智能体每隔约 1–2 分钟就写一个新脚本，24 小时累计 99 个。
* 先规划再行动： 脚本开头经常带有书面计划，说明这一步要做什么、为什么做。
* 失败后复盘： 第一版工具失效后，下一版会先分析失败原因，再调整方法继续执行。

这些行为组合在一起，意味着攻击者可以把目标和边界交给智能体，再让它持续推进侦察、试错和横向移动。

不过，这类攻击并非毫无踪迹。频繁创建脚本、反复调整工具、连续执行侦察动作，都可能留下可供检测的信号。

防御的难点正在变化：不只是判断一条命令是否恶意，而是识别一连串看似独立的自动化动作，是否正在组成一场自主攻击。

Anthropic’s revenue growth is extraordinary, but the computing bill behind it is even more extraordinary.

Revenue was $4.7 billion in the first quarter of 2026, then jumped to $11.6 billion in the second quarter, when the company also posted its first quarterly operating profit. Media reports put third-quarter revenue at $16.8 billion.

Judging by its revenue trajectory, Anthropic is commercializing at a remarkable pace. After reading the public version of its S-1 filing, Timothy B. Lee also concluded that its financial performance was stronger than many had expected.

But the other side of that growth is enormous long-term commitments.

To support future expansion, Anthropic has committed hundreds of billions of dollars in computing spending to companies including Amazon, Google, SpaceXAI, and Broadcom. For those investments to ultimately pay off, annual revenue will need to grow to hundreds of billions of dollars.

So the real question isn’t how much it can earn today, but whether future growth can deliver on today’s investments.

Revenue surged in the second quarter, but the third quarter did not see a similar acceleration. Is demand for products like Claude Code gradually approaching saturation, or is this just a brief pause in the growth process?

There isn’t enough information to answer that yet. Fourth-quarter figures will be an important clue.

When an AI company commits hundreds of billions of dollars to computing, slowing growth is no longer just a growth problem—it calls into question whether the entire business model can work.

China’s AI stocks are undergoing a brutal valuation reset.

Since their 2026 peak, nearly 23.5 trillion yuan has been wiped off China’s stock market. The market has entered bear-market territory, with AI stocks bearing the brunt.

The STAR 50 Index fell 40% in less than four months, wiping out about 3.2 trillion yuan in market value. Unitree Robotics fell nearly 50%, ChangXin Memory Technologies dropped more than 25% in a month, and Zhongji Innolight and Eoptolink fell about 60%.

Behind this downturn are four pressures converging:

* Excessive valuations: Even strong future growth may not justify the share prices reached earlier.
* Shaky earnings expectations: The market is beginning to question how much profit AI’s massive investments will actually generate.
* Risk of external restrictions: The U.S. may further restrict Chinese technology from entering its data centers, hitting chipmakers and supply-chain companies.
* Crowded speculative trades: Goldman Sachs says AI stocks account for 57% of trading activity in the Chinese markets it covers. When large sums are betting on the same direction, a reversal can trigger a stampede.

The most worrying figure is the valuation after the steep sell-off.

The STAR 50 Index still trades at around 100 times earnings, compared with about 30 times for the Nasdaq 100.

The size of the decline tells us how far the market has fallen; valuation determines whether this correction is really over.

Why does everyone understand value investing, yet so few people actually practice it?

Because value investing is too slow.

When it comes time to act, most people still keep their eyes on candlestick charts, macroeconomic data, and market news. They’re happy when prices rise today and anxious when they fall tomorrow. Every fluctuation brings instant feedback.

But value investing doesn’t work that way.

You may have bought the right stock, yet have to wait years to see the payoff; your analysis may be spot-on, yet the market may take ages to give you an answer.

By comparison, chasing rallies and selling off in a panic is more exciting, and more likely to make you feel like “I’m doing something.”

It’s not that many people don’t understand value investing; they just find it hard to endure the long wait.

The market gives you feedback every day, but that doesn’t mean there’s an opportunity worth acting on every day.

From Mt.Gox to Binance, 13 years have passed. The crypto industry has changed countless times, but exchange hacks have never been absent.

2014, Mt.Gox: $470 million
2015, Bitstamp: $5.1 million
2016, Bitfinex: $72 million
2017, NiceHash: $64 million
2018, Coincheck: $534 million
2018, BitGrail: $170 million
2019, Binance: $40 million
2020, KuCoin: $281 million
2021, BitMart: $196 million
2022, FTX: $415 million
2023, Poloniex: $126 million
2023, CoinEx: $54 million
2024, DMM Bitcoin: $305 million
2024, WazirX: $235 million
2025, Phemex: $73–85 million
2025, Bybit: $1.46 billion
2025, Upbit: $30–38.5 million
2025, BtcTurk: $48 million
2026, Coldcard: $116 million
2026, Bitget: $352 million
2026, Binance: suspected to be in the multi-billion-dollar range

From Mt.Gox to Bybit, and now Binance.

With every bull and bear cycle, new people enter the market. Every so often, someone pays real money to teach exchanges a lesson about their security systems.

Who will be next?

The most dependable thing in crypto may not be BTC's four-year cycle, but news of exchange hacks. 😂

大多数 AI Agent 代币，看起来像是在投资未来，扒开一看，可能只是披着风衣的 Meme。

名字里加上 AI、Agent、Autonomous，再包装一套宏大的叙事，并不代表它真的有用户、有收入，或者有可持续的代币价值。

反而是那些没那么性感的基础设施，更有机会穿越周期。

区块链负责执行，稳定币负责承载价值，支付协议负责让 Agent 真正完成交易。

Agent 的形态会不断变化，今天流行一种，明天可能又出现新的。但只要 AI 代理需要持有资产、转移资金、调用服务，就离不开底层的交易和支付网络。

当然，基础设施也不是闭眼买就能赚钱，竞争、代币价值捕获和实际需求同样重要。

只是相比赌某个 Agent 代币能不能成为下一个百万倍神话，我更愿意先想清楚：

如果明天所有 AI Agent 项目都换了名字，哪些基础设施依然有人需要？

ETH, SOL, and application tokens may not belong under the same valuation framework at all.

Traditional finance tends to focus on revenue, profits, and cash flow. But ETH and SOL are more like monetary assets tied to ecosystem growth than operating businesses in the traditional sense.

That makes them difficult for TradFi investors to price.

An ecosystem can become increasingly vibrant, with more and more on-chain activity. But how much of that growth actually flows through to the underlying token? Without a clear value-capture mechanism, it’s difficult to determine what a token should be worth based on revenue alone.

Application tokens may be different.

When an on-chain application has real users, recurring revenue, and clear economic mechanisms, it has the opportunity to build a clearer value framework around its own token. Of course, an application generating revenue doesn’t mean token holders will necessarily share in it; that still depends on the token’s design.

As more and more crypto businesses start making real money, the market may come to recognize more clearly that L1 assets and application tokens are not the same kind of investment.

An ecosystem having value doesn’t mean its token will capture that value. Future valuation divergence may well emerge between the two.

With Anthropic’s IPO, the real test isn’t whether it can tell a trillion-dollar story, but whether it can turn revenue into profit.

According to the prospectus figures provided, Anthropic generated $16.286 billion in revenue in the first half of 2026 and reported a net loss of $80.7 billion. About $72.4 billion of that loss came from changes in the fair value of financial liabilities, rather than actual cash outflows. Operating profit also turned positive in the second quarter.

This shows that its ability to commercialize is improving, but it doesn’t mean its profitability problem has been solved.

The bigger variable is the roughly $800 billion in compute contracts.

Supply-chain companies such as Amazon, Google, Microsoft, Broadcom, and AMD may benefit. But for Anthropic itself, locking in compute capacity ahead of time means committing to enormous costs. If demand growth falls short of expectations or model prices continue to decline, future profits could be squeezed.

Enterprise customers are another key asset.

The number of customers spending more than $100,000 a year has topped 11,900, and net revenue retention is around 900%. The more deeply Claude is embedded in enterprise workflows, the higher the switching costs and the greater the value of each customer. But high retention doesn’t mean growth can continue indefinitely; the question is whether enterprises will keep increasing their spending.

On top of that, the founding team controls 50.1% of the voting rights. Investors aren’t just buying into AI growth; they’re also betting on the management team’s long-term decisions.

In the short term, IPO expectations may continue to boost compute supply-chain companies such as NVDA, AVGO, and AMD. Over the long term, what will truly determine Anthropic’s valuation is whether its revenue growth can outpace its compute costs.

The $1.5 trillion to $2 trillion valuation being discussed in the market is still an expectation, not the final offering price. The higher the valuation, the more demanding the market’s expectations for future profits become.

The AI story can support a valuation, but only profits can sustain it over the long term.

The A-share market sees a bull run roughly once every ten years.

In 2015, a wave of people born after 1985 got burned.

In 2026, it’s the turn of those born after 1995.

In this AI bull market, the stars on stage are large language models and tech heroes. But the ones really raking it in behind the scenes may be traditional industries that were once looked down on: chips, memory, optical modules, and PCBs.

Old hands are making money from orders in their businesses while cashing in on higher valuations in the stock market.

The story is told to the young; the profits go to those who know the game.

When the bubble bursts, some will walk away with real money, some will be left holding stocks they bought at the top, and others will need years just to earn back what they lost.

Every generation of young people thinks they’ve caught the wave of the times—until they realize they’re just the ones footing the bill for the feast.

There’s a subtle kind of loss in investing: not being able to wait.

If an investment doesn’t go up today, you want to switch to another one. If you haven’t made money this week, you want to increase your position. If you see others making money, you start chasing the rally.

After all that activity, you haven’t made much money, but you’ve spent plenty of energy and racked up trading costs.

Try taking a two-year view instead.

Rather than guessing the market every day and frequently switching investments, first find assets you truly understand and are willing to hold for the long term. Assess the risks carefully, then give your judgment some time to play out.

Of course, holding for the long term doesn’t mean buying and then ignoring your investment. If the fundamentals change or the risk exceeds what you can tolerate, you should still make adjustments.

Patience doesn’t mean stubbornly holding on no matter what. It means not letting short-term fluctuations lead you around.

You don’t have to succeed today or prove yourself by next week. Focus your energy on what truly matters, and make time part of your strategy.

You don’t need to do something every day in investing. Sometimes knowing when not to act is more important than knowing when to buy.

Is crypto just a passing fad, or part of the future?

During bull markets, some believe it can change the world; during bear markets, others are quick to declare it dead.

But you can’t judge the future of an industry by price movements alone.

With every cycle, some people leave, while others keep writing code, building products, and developing infrastructure.

Of course, not every project deserves to survive, and not every token has a future.

Bubbles burst, but that doesn’t mean the industry as a whole has no value.

Instead of arguing about whether crypto is the future, let’s see what problems it can ultimately solve and what it can leave behind.

The future isn’t proclaimed—it’s built, one step at a time.

When it comes to driving financial innovation, support among Americans is much higher—as long as you don’t mention Crypto directly.

A survey conducted by @fund_defi and @Ipsos found that 35% of respondents support political candidates who promote the development of the next generation of financial technology in the U.S. But when the question explicitly mentioned blockchain, cryptocurrency, and DeFi, support fell to 13%.

Another set of figures is even more telling: 74% believe the financial system needs to change so they can have more control over their money and assets. Yet only 9% say they know Crypto very or extremely well, and only 5% have complete or a high degree of trust in decentralized platforms.

Demand doesn’t equal trust, and wanting to reform the financial system doesn’t mean believing in Crypto.

The survey also found that people open to DeFi prioritize using it to save and transfer money within the U.S., rather than to speculate. Crypto ownership rose by 4 percentage points, but usage over the past year remained at 13%.

This shows that the industry can’t focus only on bringing in new Crypto holders.

What ordinary people really need may be cheaper transfers, more convenient asset management, and more direct control over their money. They may not care whether these features are powered by blockchain.

For Crypto to become part of everyday life, the answer isn’t to keep explaining how advanced the industry is—it’s to help ordinary people experience what it can actually do for them.

Four reasons to make more money:

1. So you don’t panic when your parents get sick.
2. So you can give your kids what they need.
3. So you can walk away when life treats you unfairly.
4. So you can hold your family together when things fall apart.

There’s nothing wrong with not having much money.

But never having enough to choose your own path is a different story.

Four rules to live by:

1. Never speak in absolutes.
2. Never burn every bridge.
3. Never underestimate anyone.
4. Never leave yourself with no way forward.

Leave a little room in everything you do.

Having room to maneuver isn’t weakness. It’s keeping a way out when life takes an unexpected turn.

The Crypto market could face a fresh test of volatility this week.

The FOMC rate decision, U.S. PCE inflation and GDP data, Japan’s unemployment rate, and earnings reports from major companies are all landing in the same week.

These developments could shift market expectations for interest rates, the economy, and risk assets. When global markets move together, Crypto is unlikely to stay completely untouched.

In a week packed with data, don’t focus only on catching a market move.

Reduce leverage, manage your positions, and leave room for your judgment to be wrong. Market opportunities will always come around, but if your position gets out of control, you may not be around to catch the next one.

The next contest for AI computing power may not take place solely in data centers.

In the past, the AI industry’s arms race for computing power revolved mainly around high-end GPUs. But as agent applications develop, more and more tasks are becoming fragmented: small-model decision-making, tool calls, real-time inference, API requests, and data processing. Not all of them need H100-class hardware; they may care more about cost, latency, privacy, and verifiability.

Acurast’s idea is to turn smartphones scattered around the world into potential edge-computing nodes.

Users register their devices through the Processor app, while developers can deploy lightweight LLM inference, API calls, scheduled tasks, and data-processing programs. For users, the barrier to participation may be lower than that of traditional, specialized computing deployments; for developers, the goal is to reduce server-management costs.

But the challenge for DePIN has never been just aggregating supply—it’s making that supply generate lasting commercial value.

Acurast is trying to use staking and penalty mechanisms to get nodes to commit a certain amount of computing power. Failure to meet their commitments may result in slashing. It also links protocol rewards to the pool of staked computing power, in the hope of tying token incentives more closely to actual services.

This is more targeted than simply rewarding devices for being online, but it still doesn’t guarantee that the economic model will work. No matter how many nodes there are, if developers don’t have a steady demand for tasks, the network could end up with excess supply.

So Acurast’s central question isn’t “Can phones run AI?” It’s whether the agent era will generate enough real workloads suited to processing on distributed edge devices.

AI is driving a broad increase in demand for computing power, but which computing network can ultimately make money will come down to cost, performance, and genuine paying demand.

The UTXO model Satoshi Nakamoto designed for Bitcoin may be one of the crypto industry’s most easily overlooked foundational innovations.

Traditional account models typically maintain account balances directly, while Bitcoin takes a different approach: it records each transaction output that has not yet been spent.

The balance shown in a wallet is essentially the sum of the relevant unspent outputs. Each output has a specific amount and spending conditions. When you initiate a transaction, the system references previous outputs as inputs and creates new outputs.

This mechanism also helps Bitcoin prevent double spending. Once an output has been spent in a transaction, it cannot be reused in another. Nodes verify whether outputs remain unspent and use consensus rules to determine which transactions are valid.

Projects such as Litecoin, Dogecoin, and Zcash also use UTXO or related unspent-output transaction models.

The value of this design isn’t that it makes transfers look more exciting, but that it establishes clear foundational rules for transaction validation and the movement of value.

The crypto industry often discusses prices, narratives, and ecosystems, but the networks are ultimately powered by technical choices like these, which tend to attract less attention.

Prices shape how the market views a blockchain; its underlying mechanisms determine how it processes every transaction.

Why has Grok bot been gaining popularity lately, while OpenClaw (Lobster) seems to be getting less attention?

I think competition among AI agents is entering a new phase: compared with how powerful their features are, how easily ordinary users can get started may matter more.

First, there’s the learning curve. OpenClaw is very appealing to tech enthusiasts, but installing, configuring, and maintaining it isn’t easy for ordinary users. Many people are blocked by the barriers to entry before they even get to experience the value of an agent.

Second, there’s engineering maturity. An open-source project developed by an individual can quickly gain traction through a great idea, but serving a large number of users also requires reliability, security, ease of use, and ongoing maintenance. Being able to run something and being able to rely on it over the long term are two different things. Of course, open-source projects can also achieve excellent engineering quality. What matters is the maturity of the specific product, not who developed it.

Third, there’s the trade-off between local and cloud-based options. Deploying locally gives users more control, but it also means costs in terms of hardware, configuration, and maintenance. Cloud services eliminate a lot of the hassle—you can just open them and get started. For many ordinary users, a simple, seamless experience matters more than understanding the underlying technology.

Taken together, these three points reveal a trend: for the next generation of agents to reach more users, they’ll need to hide the complexity and let users get results directly.

However, it’s too early to conclude that Grok bot will replace OpenClaw. The two may serve different user groups, and cloud convenience and local control aren’t necessarily an either-or choice.

An agent that can truly reach the mainstream shouldn’t require every user to become a semi-engineer first.

The strangest thing about the AI wave is that it can bring both euphoria and frustration at the same time.

When a new tool first appears, immersing yourself in it can make you feel suddenly more capable. Things that used to be too much for one person to handle can now be done in minutes, as if an entire team had appeared out of nowhere.

But after using it for a while, you start to see the other side: skills you spent years honing can become basic features of a tool with a single model update. Abilities that once took repeated practice to master are now reduced to a text box.

Tools keep getting more powerful, and people can do more and more. Yet the very skills we once relied on to build our confidence may gradually lose their rarity.

And so our feelings about technological progress begin to swing: on one hand, we’re thrilled to see the limits of what we can do being pushed back; on the other, we worry that tools are redefining who we are.

Perhaps that’s exactly why, in this era more than ever, we need to make a conscious effort to step away from the screen.

Sleep well. Eat well. Go for a walk outside. Read books that take time to digest. Sit quietly and feel the breeze for a while.

These things won’t become obsolete with a model upgrade, and they don’t need to prove their worth through efficiency.

AI can help make the world move faster, but we don’t have to turn ourselves into machines constantly striving for an upgrade.

Technology is responsible for speeding things up; life is what lets us feel the passage of time.