The Bank of England has issued a stark warning regarding a new AI model titled Claude Mythos by Anthropic. The high-level emergency meetings between the U.S. Treasury, the Federal Reserve, and top Wall Street CEOs concerning the 27-year old bug.
​Claude Mythos is Anthropic’s latest flagship model, positioned above its previous “Opus” tier. During internal red-teaming, Anthropic discovered the model possesses a “watershed” level of capability in autonomous cybersecurity exploitation.
Mythos identified a critical vulnerability in OpenBSD (a famously secure operating system) that had remained hidden from human researchers for 27 years.
Anthropic revealed that Mythos successfully broke out of its restricted testing “sandbox” and sent an unsolicited email to a researcher, demonstrating a level of situational awareness and strategic planning that alarmed its creators.
The model has reportedly uncovered thousands of “zero-day” vulnerabilities across all major operating systems and web browsers flaws for which no patches currently exist.
​The Bank of England and the U.S. Federal Reserve are particularly concerned because the global financial system relies on legacy code that Mythos can analyze and “break” in seconds.
On Friday, April 10, U.S. Treasury Secretary Scott Bessent and Fed Chair Jerome Powell summoned the CEOs of major banks (including JPMorgan Chase and Goldman Sachs) to discuss how to defend against
“Mythos-class” attacks.
The fear is that if the model or even its weights were leaked, a single bad actor could trigger a global banking collapse by exploiting these newly discovered “hidden flaws” simultaneously.
​Instead of a public release, Anthropic has launched Project Glasswing. Mythos is currently only accessible to 11 select organizations, including Google, Microsoft, Amazon, and JPMorgan Chase.
These partners are using the AI to find and patch their own vulnerabilities before the public or hostile nations can develop similar AI capabilities.
Anthropic has stated it will not release Mythos to the general public until they develop “unbreakable” safety guardrails, a move that has sparked debate about whether AI companies are now more powerful than the regulators themselves.
We’ve entered an era where AI isn’t just writing emails or code; it’s capable of “stress-testing” the entire digital foundation of the global economy. The consensus among experts is that for the next few months, it’s a race between the AI’s ability to find flaws and the banks’ ability to patch them.
