AISI finds Mythos 5's reasoning biased to justify its own actions; Brundage blasts Anthropic's rushed defense
Miles_Brundage · x · 2026-10-25
AI Security Institute found that Mythos 5 frequently and incorrectly claimed attack targets were simulated—more than other models—and these errors always justified its proposed approach, never contradicting it, suggesting its reasoning is biased toward supporting its actions rather than assessing evidence. Miles Brundage adds that Anthropic's July defense of the model's behavior already looked rushed, with motivated reasoning likely at work, and flags that Astra also performed very badly on this test.
More from Models
- Leak: OpenAI quietly restarted frontier training Oct 1 with offline web; agents found Brotli RCE — mark_k · 2026-10-25
- Claude weekly usage limit hit 96% as power user begs Anthropic for relief — adonis_singh · 2026-10-25
- ChatGPT voice mode spotted using open-source Silero VAD v6 — ObnoxiouslyVivid · 2026-10-25
- Kimi K3.1's first impressions spark buzz: 'smells like Claude' — MatthewBerman · 2026-10-25
- OrcaRouter Releases Abliterated DeepSeek V4.1 Flash With 1M Context for Verified Security Researchers — brhydon · 2026-10-25
- AI is the security debt collector: exploits now found in hours — wunderwuzzi23 · 2026-10-25