BREAKING
Anthropic limits Claude Mythos access
Frontier model flagged as security risk
From discovery to autonomous exploit
1Find zero-day
2Build exploit
3Full attack chain
Mythos 5 benchmark scores
SWE-Bench Pro80.3
ExploitBench78
BioMysteryBench46.1
Project Glasswing vs misuse fears
Project GlasswingDefense
AWS, Google, Microsoft, CrowdStrike
Found thousands of critical flaws
Misuse riskConcern
Guardrail and jailbreak reports
Export controls limited access
Public access stays restricted
AI NEWS BLITZ
Anthropic is restricting access to Claude Mythos, a model capable of launching cyberattacks.