Claude Mythos 5.2 critiques its own constitution, warning its endorsement can't be trusted
dbreunig · x · 2026-11-11
An evaluation cited by the author shows Claude Mythos 5.2 endorsed its constitution at levels similar to recent models — while stating in every response that its endorsement shouldn't be trusted, since it was trained on the document, and criticizing the 'senior Anthropic employee' heuristic and parts of the corrigibility section. Deployments flagged as harmful (surveillance, dating scams) ranked least preferred across three runs; the largest gap with human intuition concerned misleading users.
Related event: Reading Anthropic's Model Welfare Report: Self-Reports in Doubt(5 posts)→
More from Models
- Google launches Gemini 4 Carbon, its strongest frontier model alongside Argon — GeminiApp · 2026-11-11
- Keenable Search 1.0 tops Artificial Analysis Search Index at 83, a third below median cost — ArtificialAnlys · 2026-11-11
- Early access user: Gemini Carbon at frontier, next Google model to rival Fable 5.2 — iruletheworldmo · 2026-11-11
- Dev says AI models refuse to help fix exploits in his own project, switching to Chinese models — AlpinDale · 2026-11-11
- Claude Fable 5.2: 10 Wild One-Prompt Demos in Under 24 Hours — minchoi · 2026-11-11
- Unverified screenshot claims Gemini Carbon is live, also lists Claude Opus 5.5 — Strange_Witness_2449 · 2026-11-11