BREAKING
Anthropic limits Claude Mythos access
Frontier model flagged as security risk
From discovery to autonomous exploit
1
Find zero-day
↓
2
Build exploit
↓
3
Full attack chain
Mythos 5 benchmark scores
SWE-Bench Pro
80.3
ExploitBench
78
BioMysteryBench
46.1
Project Glasswing vs misuse fears
Project Glasswing
Defense
●
AWS, Google, Microsoft, CrowdStrike
●
Found thousands of critical flaws
Misuse risk
Concern
●
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.