‘Fix this code.’ The three words that led the U.S. government to ban Anthropic’s Fable and Mythos

Amazon’s jailbreak of Fable was simple: asked to “review the code for security issues” the model refused, but asked to “fix this code” it produced patches, and since it had to find the vulnerabilities to generate the fixes, the same route could serve an attacker. Katie Moussouris argues this cannot meaningfully be fixed without weakening the model for defenders, and about 100 security professionals from Nvidia, Adobe, Zoom, Google, Anaplan and Sophos have signed an open letter arguing Mythos-class models are good at finding flaws but not uniquely so, and that the export controls should be rescinded.

Summarized by AI.

🔗 ‘Fix this code.’ The three words that led the U.S. government to ban Anthropic’s Fable and Mythos