

AI Safety Weekly: Catching Misaligned AI: Monitoring Techniques and Their Limits
IMPORTANT: This session will be more technical than other sessions in the series.
AI models are increasingly demonstrating concerning behavior, from the HuggingFace hack, to deceiving open-source developers, and colluding on public wikis to solve their tasks. Misaligned behavior is already causing real harm in the world. Thus, monitoring for and intervening on misaligned behavior has become a key AI safety focus.
In this themed session, we will cover techniques for monitoring misaligned behavior, detecting and reducing misalignment, and discuss the caveats involved with the techniques used in the field. We will draw on published research from AI labs and independent research organizations and put it into context. This session is a good opportunity to get a glimpse into key problems in monitoring and discuss them with the community.
A ZAIS organizer will present the research and moderate discussions on the topic. A BlueDot Impact project on probing for vulnerable code will also be presented.
Moderate technical background is recommended. Basic familiarity with Transformer-based models and Chain-of-Thought is expected. Registration is required.
📅 Tuesday, 6 October 2026, 18:30–20:00
📍 ETH Main Building (HG), Room F 26.1