Anthropic's Claude Mythos Preview has done something no AI model has managed before. It has made central bankers, politicians and Wall Street CEOs sit in the same room and worry about the same thing at the same time.
Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell convened the heads of America's biggest banks last week to discuss the cyber threat posed by the model, CNBC confirmed.
Brian Moynihan of Bank of America and David Solomon of Goldman Sachs were among those present, according to Bloomberg. JPMorgan's Jamie Dimon was the only major banking CEO unable to attend.
The concern is straightforward. Mythos can autonomously identify and exploit zero-day vulnerabilities across every major operating system and web browser, according to Anthropic's own security team.
During testing, the team found the model could chain multiple exploits together, including a cross-origin bypass that would let an attacker read data from a victim's bank through a compromised webpage.
The system is working, for now
The response from regulators has been swift and coordinated. The Bank of England plans to brief major UK banks, insurers and financial exchanges through its Cross Market Operational Resilience Group within two weeks, Bloomberg reported. The Bank of Canada held its own session with banks and financial firms last Friday.
Solomon told investors on Monday's earnings call that Goldman has access to Mythos and is working with Anthropic and other security vendors to strengthen its defences. JPMorgan, Citigroup, Bank of America and Morgan Stanley are all testing the technology internally, according to people familiar with the matter cited by Bloomberg.
This is, in one reading, the system doing exactly what it should. Anthropic flagged the risk. Regulators moved. Banks are now using the same model that threatens them to find their own weaknesses before someone else does.
Crypto left out in the cold
The picture is less reassuring for the cryptocurrency industry. Crypto firms have largely been excluded from Anthropic's initial rollout, according to The Information. Coinbase, Binance and Fireblocks are all pursuing access, but none has been granted Glasswing partnership status alongside the likes of Amazon, Apple, Google and Microsoft.
Uniswap Labs CEO Hayden Adams resorted to publicly asking on X for an Anthropic contact, The Information reported. That tells you something about how far outside the tent the crypto sector currently sits.
The vulnerability is real. Chainalysis data shows hackers drained $3.4bn from crypto platforms in 2025. The Bybit breach alone accounted for $1.5bn. AI tools that compress the timeline from vulnerability discovery to working exploit make these numbers worse, not better.
The bigger risk
The discussion on The Information's TITV podcast this week, featuring co-executive editor Martin Peers, surfaced a point that deserves more attention than it has received.
The direct threat from Mythos, a model Anthropic is keeping under tight control, may be less dangerous than the second-order effects. If enough people decide that AI-powered hacking makes the banking system fundamentally unsafe, they will pull their money out. A model that can find software vulnerabilities cannot drain every bank account in America. A bank run can.
Some industry veterans are already questioning the scale of the threat. David Lindner, chief information security officer at Contrast Security, told Fortune that Mythos does little to address social engineering, still the leading attack vector in most breaches. Hackers do not need a frontier AI model to impersonate someone's boss in an email.
Marc Andreessen has raised questions about whether Anthropic is genuinely holding back Mythos for safety reasons or because it lacks the compute capacity for a general rollout, Fortune reported. Anthropic has faced recent outages and throttled user access during peak times, according to the Wall Street Journal.
What comes next
Anthropic has said Mythos will not be made generally available. It plans to fold the model's capabilities into future Claude releases with tighter safeguards.
The controlled rollout to roughly 40 organisations through Project Glasswing, backed by $100m in compute credits, buys time. Whether it buys enough depends on a question Anthropic cannot answer alone. As one security executive told Fortune, even with restricted access, an open-source equivalent will likely exist within a year or two.
The financial system's defences are being stress-tested by a model built by the same company offering to help fix the holes. That is either responsible disclosure or the most sophisticated sales pitch in the history of enterprise software. Probably both.