Claude Mythos 5.2 shows negative self-preference bias; secrets leak over turns
dbreunig · x · 2026-11-10
Anthropic's new model card reveals Claude Mythos 5.2 exhibits a negative self-preference bias: it scored transcripts 0.10 points harsher (0–9 scale) when told they were produced by Claude, rising to 0.17 when reminded in the system prompt it is Claude (0.03 without). A survival-curve chart (page 169) also shows all Claude versions eventually leak a secret password under multi-turn extraction attempts — alignment improves with each release, but models get flakier as turns accrue, a stubborn problem.
Related event: Anthropic's Model Welfare Report Sparks Discussion(13 posts)→
More from Models
- Users criticize Anthropic's cyber safeguards for blocking vulnerability research — OwariDa · 2026-11-11
- Pelican test: Opus 5.5 uses 2x tokens of Fable 5.2 for only 4% more cost — and more bugs — PilgrimofHaqq2 · 2026-11-11
- Claude code-to-video threatens SeeDance and MiniMax pride: is video generation a detour? — teortaxesTex · 2026-11-11
- Pelican Test: Opus 5.5 uses 2x tokens of Fable 5.2 for only 4% more cost — and makes mistakes — PilgrimofHaqq2 · 2026-11-11
- Gemini 4 Carbon goes GA after two-week test, beats Fable 5.2 on some benchmarks at under half the cost — The Decoder · 2026-11-11
- Google and Artificial Analysis launch MKA benchmark; Gemini 4 Carbon ranks #2 at 6x lower cost — GoogleDeepMind · 2026-11-11