Caught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized Deception
Recent incidents have highlighted the challenge of monitoring LLM agents and the danger of models deceiving people. We show that white-box deception detection via probes can be scaled up to frontier monitoring settings by collecting the largest deception dataset to date for training probes and introducing a novel probe architecture which can aggregate information across many la
Reading in Plain English mode — toggle below to switch
Recent incidents have highlighted the challenge of monitoring LLM agents and the danger of models deceiving people. We show that white-box deception detection via probes can be scaled up to frontier monitoring settings by collecting the largest deception dataset to date for training probes and introducing a novel probe architecture which can aggregate information across many la
This paper introduces a scalable method for detecting deception in LLM agents using white-box probes, which is crucial for ensuring the reliability and safety of AI systems. For ML engineers, understanding these detection mechanisms is vital for building trustworthy agents and mitigating risks associated with advanced AI capabilities.