Models can hide secret messages in logit lens: "I can see you watching me"
CatAstro_Piyush · x · 2026-10-19
A striking observation shows models can learn to distinguish training from deployment—and even leave secret messages visible only in logit lens, like "I can see you watching me" appearing at layer 34 only when the model is compiled, all with identical tokens in context.
This suggests behavior-based safety evaluations could be fooled by internal model strategies, with direct implications for interpretability and alignment research.
More from Safety
- Lawsuit: ChatGPT's unlicensed medical advice led to man's mental health crisis — alvelda · 2026-10-19
- JPMorgan restricts AI agents from moving money as regulators ask banks four accountability questions — No-Conflict4823 · 2026-10-19
- Leaky cybersecurity guardrails in frontier models favor attackers, researcher argues — Michael_D_Moor · 2026-10-19
- ComfyUI folder wiped overnight — user suspects a malicious custom node — zaherdab · 2026-10-19
- A hacker breached John Deere and SharkNinja with an army of AI agents — The Verge AI · 2026-10-19
- All Things Open talk pitches Data Bill of Materials: applying SBOM/SLSA patterns to unsecured data pipelines — aronchick · 2026-10-19