Model Forensics: Investigating Whether Concerning Behavior Reflects Misalignment
Singh, Kroiz, Rajamanoharan, Nanda · 2026 · arXiv:2606.26071
Detecting concerning behaviour is not the same as establishing misalignment. Proposes a two-step protocol: read the chain of thought to form hypotheses, then edit the environment to test them.
misalignmentdeceptionchain-of-thoughtevalsagentsauditing
Discussed in
- BASIS 09, Reading Group 9: Model Forensics · 25 Aug 2026