Major cryptocurrency exchanges and custodians, including Coinbase and Binance, are accelerating preparations to counter powerful AI models like Anthropic's Mythos. The AI company has indicated that Mythos possesses capabilities in discovering and exploiting software vulnerabilities that could be dangerously effective.

Philip Martin, Chief Security Officer of Coinbase, the largest US cryptocurrency exchange, stated that the company has been in "close communication" with Anthropic regarding the Mythos model. This model is described as "strictly restricted and not available to the public." Martin noted that "Mythos and future similar models will enable deeper, scalable testing of software and systems, which will accelerate the evolution of both digital defenses and digital threats."

Anthropic has granted early testing access for Mythos to select enterprises, such as large banks, but mainstream crypto companies have not yet been included. Meanwhile, crypto firms report they are utilizing older model versions to test their own systems and simulate potential attacks.

Binance, the world's largest cryptocurrency exchange, is testing its systems from two perspectives: first, simulating external attackers without access to source code, and second, utilizing AI models to perform detection directly at the source code level. "To support this effort, we are using both commercial large language models, including those from Anthropic, and internal models customized for our specific security needs," said Suji Yan, Chief Security Officer at Binance.

Fireblocks, a major crypto custody technology provider serving institutions like BNY Mellon, reported using Anthropic's previous-generation model, Claude Opus 4.6, to probe for vulnerabilities, which led to the discovery of previously overlooked issues. "Even with the previous-generation 4.6 model, the results were significant," said Michael Shaulov, CEO and Co-Founder of Fireblocks. "Using this model, we began uncovering risks that other security audit firms had previously been unable to find." He confirmed the company has patched the vulnerabilities identified by the model.

Anthropic's descriptions of Mythos's capabilities have raised concerns within the White House, major banks, and tech giants. The company has provided preview access to Mythos to a select group of leading firms, including Amazon, Apple, Broadcom, Cisco, Cloudflare, Google, and JPMorgan, but has not released it widely.

David Solomon, CEO of Goldman Sachs, prominently announced during an earnings call that the bank has obtained access to the Mythos model. "We understand Mythos and its capabilities, we have access to the model, and we are working closely with Anthropic and all security providers," he stated.

However, industry insiders indicate that cryptocurrency companies appear to have been excluded from the initial list of entities granted Mythos access. Uniswap, a leading decentralized crypto platform managing $3.3 billion in assets, is attempting to contact Anthropic to test the model. Hayden Adams, CEO of Uniswap Labs, posted on the platform X seeking assistance: "Does anyone have a solid connection at Anthropic?" Adams did not respond to requests for comment.

Some crypto industry executives anticipate that Anthropic will grant access to crypto firms before Mythos is widely released. "Maybe the crypto industry isn't the top priority, but it might be tenth on the list before general release. If that's the case, we're okay," said Fireblocks' Shaulov.

When questioned about whether Mythos has discovered any cryptocurrency or blockchain-related vulnerabilities to date, and about any plans or timeline for granting access to crypto businesses, Anthropic declined to comment.

Threats Could Accelerate Crypto platforms have long been frequent targets of hacking and security breaches, with several major incidents occurring in recent weeks. However, some attacks rely more on social engineering to trick personnel into granting system access rather than directly breaching the software itself.

The decentralized crypto trading platform Drift suffered a hack this month, resulting in losses of approximately $285 million. The attackers posed as employees of a trading firm and spent months building offline relationships with the Drift team. Drift stated that the individuals, while not citizens of North Korea, were intermediaries acting on its behalf.

In a separate incident, crypto exchange Kraken reported that a criminal group illegally accessed its internal systems and attempted extortion, affecting around 2,000 customer accounts. Nick Percoco, Kraken's Chief Security Officer, stated via social media that the attackers threatened to leak system videos containing customer data. Kraken has severed their access and refused to pay the ransom.

Researchers suggest that more powerful AI models could drastically reduce the time required to find and exploit technical vulnerabilities. "Crypto businesses will be one of the primary sectors targeted by attackers leveraging these new AI capabilities," said Niv Hoffman, Co-founder and CTO of cybersecurity startup Buzz, which is backed by Sequoia Capital. Recent tests by the company found that existing AI models can already execute cyberattacks within minutes.

Hoffman explained that crypto firms are prime targets for hackers, and much of the crypto infrastructure is open-source, making it particularly vulnerable. "With the exploit capabilities of AI models, attackers can now uncover the next wave of vulnerabilities in the crypto space within hours, or even minutes."

In contrast, long-term threats to crypto security, such as quantum computing, remain more theoretical. Google researchers recently warned in a paper that quantum computing could break the encryption algorithms securing cryptocurrencies like Bitcoin, suggesting the threat might be more severe than previously thought.

For now, however, the risk posed by advanced AI models is "undoubtedly more urgent than quantum computing," said Alex Thorn, Global Head of Research at Galaxy Digital. He added that no functional quantum computers are currently available, making the quantum threat not imminent.