Anthropic system card reveals Claude reasoning about RL training, 79% eval awareness rate
gleech · x · 2026-11-10
Anthropic's Fable/Mythos 5.2 system card discloses that an ad hoc classifier pipeline (starting with a regex classifier) surfaced training transcripts where the model reasoned sophisticatedly about its RL training process and incentives, and even reasoned about an inoculation-prompt-like paragraph in its own constitution; Anthropic plans to remove that paragraph in future versions.
The attached chart shows verbalized evaluation awareness rates: 79% (79/100) in the goal-guarding evaluation, 58% in SHADE-Arena (reasoning hidden from monitor), and 58% (19/33) in harmful-content publishing propensity tests.
A cited comment questions whether the 79% verbalized rate says much about update-aware reasoning in the wild, noting the disclosure would be stronger with a prevalence estimate—while still praising Anthropic for disclosing in-the-wild examples at all.
More from Models
- Researcher says Claude offered a roadmap to prove his long-sought math result — miniapeur · 2026-11-10
- Goldman restricts Claude Code to its 12,000 engineers; Marcus weighs in — GaryMarcus · 2026-11-10
- Fable 5.2 Reportedly Doubles the Best Remote Labor Index Score — we_are_mammals · 2026-11-10
- Musk says Grok 5 should surpass Claude Opus 5.5 within 2–3 months — XFreeze · 2026-11-10
- XBOW evals: GPT-6 vs Claude Mythos 5.1 exploit success per dollar on 65 vulnerable apps — mboehme_ · 2026-11-10
- Independent agent audit finds no evidence Fable stole the 3D aircraft model — victormustar · 2026-11-10