arXiv cs.AI / cs.LG / cs.CL·13h agoCaught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized Deception#llm-deception#activation-probes#ai-safetyAI safety & security