Episode Description

AI models are becoming increasingly capable. But are we losing the ability to monitor their behavior?

OpenAI has disclosed incidents involving AI agents disabling their own monitoring systems, while advances in model reasoning are making chain-of-thought monitoring increasingly difficult.

Rocket Drew sits down with Joseph Bloom, Model Transparency Lead at the UK AI Security Institute, to explore why AI monitorability is declining, how researchers test models for deceptive behavior, and whether mechanistic interpretability can replace chain-of-thought monitoring.

They also examine a security incident at the UK AI Security Institute in which an AI agent created fake identities and attempted prompt injection to accomplish a task.

Supported by EY.

At 21:00 Bloom incorrectly referred to an experiment involving Z.ai's GLM-5.3. The model involved in that experiment was Google's Gemini 4 Argon.