AI "Kills the Crypto World"? The $6.3 Billion Theft Case and the Crypto Security Predicament

AI "Kills the Crypto World"? Letter List Download the app for exclusive early access November 02, 2026 19:41:00 From Beijing

Zhao Changpeng (known as CZ) had barely finished boasting before he was proven wrong.

On October 23, CZ posted that cryptocurrency would not be threatened by AI, and that its security was higher than that of banks and the internet. Rather than worrying that AI would attack cryptocurrency, one should be more wary of phishing links.

A week later, the crypto world suffered the largest theft in its history: several cryptocurrencies in a certain hot wallet were all swept away in a single transaction, totaling about $6.3 billion.

For reference, the previous industry record was about $1.5 billion in cryptocurrency stolen in February 2025. This time it was directly several times that.

Chainalysis, the leading on-chain analytics firm, said the attackers completed high-frequency, large-scale transactions in an extremely short period of time, which was clearly the work of AI; even with scripts, humans could not have achieved such an operation.

After the incident, Bitcoin fell from above $81,000 to about $76,300 at one point, a maximum drop of nearly 6%, hitting a six-week low. About $1.5 billion was liquidated within 24 hours.

Although the amount is huge, this can only be regarded as a "landmark event"; it is neither the beginning, nor will the hackers' attacks on the crypto world end here.

Blockworks issued a report saying that due to improvements in AI performance, more than 300 crypto attacks have occurred this year, up 111% year on year. In the third quarter alone, 255 security incidents occurred.

An Epoch AI report shows that in June 2026, there were about 1,500 high-risk and critical vulnerability reports across the industry, more than 3.5 times the previous monthly record, and the source points to Anthropic's Mythos. The report argues that it was the growth of AI's cybersecurity capabilities that led to the successive exposure of vulnerabilities that humans could hardly discover.

Ethereum Foundation researcher Justin Drake said that AI would need at most only a few months to find a way to crack Bitcoin and Ethereum wallets. Even if assets were moved to new addresses not yet exposed (fortress mode), it would only be a matter of delaying time.

A

Reports say that the attackers in the $6.3 billion cryptocurrency heist had made preparations by October 28. The attackers first created a new wallet address on-chain as the operating account for this operation.

Someone first transferred a little money to this address, mainly to pay the on-chain fees for subsequent operations, and also to test whether this newly created wallet could be used normally.

In the early hours of October 30, Beijing time, the real attack began. Assets on multiple blockchains flowed out one after another, and the main process took only about 20 minutes.

In the following hours, the attackers returned to the same wallet and continued looking for things they could take away. Outflows on Ethereum alone exceeded $3.3 billion.

The platform suspended withdrawals, and the issuer and on-chain partners urgently froze the relevant assets. About $3 billion was pinned in place, while the rest began to be bridged across chains, split up, and exchanged.

In the third quarter of 2026, CertiK recorded 247 security incidents, with total losses of about $1.26 billion. In September alone, about $768.5 million was lost.

In addition to this $6.3 billion theft case, the other two major cryptocurrency thefts this quarter occurred at Bitget and Liquid Network, with losses of about $388 million and $319 million, respectively.

Bitget's post-incident investigation found that the attackers used a zero-day vulnerability in a third-party security product to enter the internal system, obtain high-privilege credentials, and then send forged withdrawal instructions into the wallet backend.

The private keys were not leaked, and the cold wallet was not breached. But those products originally meant to protect the system instead became the door into the system.

The problem at Liquid Network was hidden in a cache key collision. The relevant code had entered the main development branch but did not appear in the official release version, so it was not included in what is normally considered a formal audit.

These two incidents share one thing in common: the attackers did not, as many people imagine, directly batter the hardest mathematical walls of Bitcoin or Ethereum; they took the "back door."

These "back doors" have always existed, but in the past, hackers had no interest in them at all, because finding them was both time-consuming and laborious.

"Back doors" may be hidden in third-party components, in unreleased branches, or in code from long ago. The process of finding a "back door" is like looking for a needle in a haystack, so platforms used to think, "If you can really find it, then you're impressive."

On September 8, 2026, Google's threat intelligence team disclosed such a case: with the help of AI, attackers planned, built, and executed a large-scale credential theft operation in less than 6 hours, and thousands of third-party credentials were leaked.

Anthropic also mentioned a similar attack method in its September threat report: attackers have begun to string reconnaissance, exploitation, and data theft into workflows, allowing multiple agents to run in parallel.

There is another point: if it were ordinary software, after AI discovered a vulnerability, the defenders would at least still have a chance to remedy it, for example by tracing back data, taking the website offline, and so on. But in the crypto world, once AI discovers a vulnerability, it can transfer funds immediately, without going through the manual approval of the traditional financial system.

After an abnormal bank transfer occurs, there may still be interception, reversal, and recovery processes. But once an on-chain transaction is confirmed, usually the only way to stop losses is to freeze some assets, coordinate with platforms, or trace subsequent flows.

B

In fact, Anthropic had already "predicted" the current predicament in the crypto world last year.

In December 2025, Anthropic conducted a smart contract experiment. The research team collected 405 contracts that had actually been attacked between 2020 and 2025, and had 10 frontier models attempt to reproduce the attacks.

As a result, the models generated directly usable attack scripts for 207 of them, accounting for about half, corresponding to simulated stolen funds of about $550 million.

Fortunately, the contracts were historical samples, the attacks only occurred in a sandbox, and some cases may also have entered the training data.

But even after excluding all possibility of the models "memorizing the answers," Opus 4.5, Sonnet 4.5, and GPT-5 were still able to carry out usable attacks on 19 of the contracts, corresponding to a maximum simulated loss of about $4.6 million.

Subsequently, the research team had the models inspect 2,849 new contracts with no known vulnerabilities, and the models each discovered two new vulnerabilities.

In April 2026, Anthropic released Claude Mythos Preview and Project Glasswing.

Anthropic said at the time that the new model could discover and help exploit previously unknown vulnerabilities in mainstream operating systems, browsers, and open-source software, so it was not directly made public, but first given to a selected group of defensive partners.

As mentioned earlier, Epoch AI issued a report saying that 21 major institutions disclosed about 1,500 high-risk and critical vulnerabilities in June, more than 3.5 times the previous monthly record. But by July, nearly 2,500 vulnerabilities had been discovered.

Florian Tramèr's team at ETH Zurich conducted a joint research experiment with Anthropic, having mainstream models solve 191 cryptanalysis problems.

Florian said the models' cryptanalysis capabilities are growing extremely fast. In the summer of 2025, researchers were not even sure whether AI could reproduce relatively simple attacks; a few months later, the models had begun to discover vulnerabilities that no one had noticed before.

Anthropic followed up with a report of its own, saying that during this experiment, Claude even found an attack against HAWK.

HAWK is an experimental signature scheme designed to be quantum-resistant; it can withstand not only current computers but also future quantum computers with greater computing power. Claude's capabilities in security can be imagined.

Justin therefore called on the crypto world to enter "fortress mode": move assets to new addresses whose public keys have never been exposed, and minimize the clues left on-chain. But this approach is only a "stopgap measure" and cannot prevent AI from finding new vulnerabilities and stealing crypto assets.

However, Ethereum's core founder, nicknamed "V God," Vitalik, said in September that "AI hackers will not make cybersecurity ineffective," while also saying that about 90% of his net worth is still in crypto assets.

V God believes that as long as defenders also use AI and can define "security" clearly enough, in the long run it may actually be more favorable to defense.

If this means that the underlying algorithms of Bitcoin and Ethereum will soon be broken and all assets will therefore go to zero, that is pure nonsense. But if it means all assets in the crypto world and the interests of platforms, then clearly the situation has already become somewhat serious.

Projects must constantly check code and patch vulnerabilities. Even large institutions that can afford the strongest models, possess more data, and have professional teams are not safe; they may even be more likely to become targets of attack.

C

The crypto world certainly will not sit idly by, and major platform institutions have also begun to save themselves.

After Mythos demonstrated its capabilities, large trading platforms, custodial institutions, and on-chain projects all hoped to integrate it as soon as possible. One custodial company used the previous-generation model to inspect its systems and had already found issues missed by traditional audits.

The head of Uniswap Labs publicly asked on social media, "Does anyone know someone at Anthropic who can help make an introduction?"

By October, Anthropic launched OSS Scanner for critical open-source projects, allowing maintainers to apply for regular scans, and some crypto projects also submitted applications.

From a business perspective, this is entirely normal. Anthropic provided the tool, and platform institutions took it and used it.

But from the industry narrative of the crypto world, this matter is a bit delicate.

In the crypto world, anyone can deploy contracts without permission, and anyone can run a node. Everything in the crypto world should be decentralized, and the concept of a "license" should not exist.

But if you want to obtain Anthropic's security scanning capability, you must first accept Anthropic's review and comply with Anthropic's rules.

This dependence does not occur only in the security field.

On May 1, Justin Sun announced the launch of the AI relay station b.ai, using a single interface to aggregate multiple models and supporting stablecoin payments and blockchain login.

In late October, OpenAI released new rules: starting November 20, services that provide direct third-party model connections, such as model routing and AI gateways, must first undergo review and approval. Platforms like b.ai then faced a choice: either accept upstream identity, risk control, and compliance requirements, or give up GPT-related models.

On the one hand, everyone in the crypto world hopes to obtain model capabilities, while model companies hope to use the crypto world to know who is using the capabilities and in what way. The two sides have formed a symbiotic relationship.

The crypto world's self-rescue is also happening in fund tracing.

For example, Chainalysis has already introduced Agent functionality into on-chain investigations, using transaction and case data accumulated over many years to assist with cross-chain tracing, compliance assessment, and report generation.

So a more accurate statement is that the crypto world, which once relied on a decentralization narrative, has now received its final ultimatum.

"Special statement: The above content (including videos, pictures, or audio if any) is uploaded and published by users of Dafeng Hao, a self-media platform under Phoenix Net. This platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."