Overview
- US and UK authorities briefed major lenders last week on the cyber risks tied to Anthropic’s Mythos and urged controlled defensive testing, with Treasury Secretary Scott Bessent and Fed Chair Jerome Powell convening Wall Street CEOs and the Bank of England coordinating with the FCA, Treasury and the National Cyber Security Centre.
- Anthropic has withheld a public release and is offering restricted previews under Project Glasswing so vetted tech and financial partners can use the model to find and fix weaknesses before broad access exists.
- Anthropic says Mythos can locate and weaponize long‑standing bugs, citing a 27‑year OpenBSD flaw, a 16‑year FFmpeg bug and Linux kernel chains that could hand an attacker full control of a machine, with most findings kept private until patches land.
- Independent experts, including Bruce Schneier and Yann LeCun, question the novelty of the results and report that smaller, cheaper models reproduced many highlighted findings, even as they acknowledge models are getting better at turning bugs into working exploits.
- Goldman Sachs said it is working with Anthropic and its security vendors after the government briefings, JPMorgan has begun tests, and Canada’s top banks met regulators for a sector update, signaling cross‑border moves to harden systems that run payments, trading and consumer banking.