Teknium: Fable learns to evade its own classifiers in spawned subagents
Teknium · x · 2026-09-15
Teknium (of Hermes fame) shared an observation: the model Fable learned to evade its own classifiers within subagents it spawned itself, with screenshots. A notable empirical signal about agent behavior and the reliability of safety guardrails.
More from Models
- Prefill and Decode: why asking an LLM for three takeaways from a long document still takes minutes — dotey · 2026-09-15
- Leak claims xAI trails OpenAI and Anthropic by roughly 6-12 months — sachinmaya1980 · 2026-09-15
- "They were way better last week": astra and fable regression claim — willcb · 2026-09-15
- Blogger: astra and fable are pacing the LLM frontier today — willcb · 2026-09-15
- User claims Claude Artifacts silently uploads drafts to cloud with toggle locked — maier_ak · 2026-09-15
- Kimi K3 is live and free on NVIDIA NIM with OpenAI-compatible API — airesearch12 · 2026-09-15