Detecting and responding to AI failures | IASEAI '26

How can we detect, evaluate, and respond to failures in advanced AI systems? This breakout session from IASEAI 2026 focuses on technical approaches to identifying and mitigating risks in frontier AI models — from benchmarking agent behavior under legal constraints to uncovering hidden vulnerabilities, bias, and misalignment in large-scale systems. Across two segments, speakers explore methods for evaluating model behavior, understanding failure modes, and designing more robust safeguards — alongside broader questions of coordination, policy gaps, and trustworthy documentation. Tiwai Mhundwa — EU-Agent-Bench: Measuring Illegal Behavior of LLM Agents Under EU Law Jacob Haimes — Sandbag Detection through Noise Injection Victoria Krakovna — Evaluating Frontier Models for Stealth and Situational Awareness Qinyuan Wu — Rote Learning Considered Useful: Generalizing over Memorized Data in LLMs Carolyn Chong — Detecting and Reasoning about Bias in Multimodal Content Daniel Herrmann — A Decision-Theoretic Approach for Managing Misalignment Punya Syon Pandey — Accidental Vulnerability: Factors in Fine-Tuning that Shift Model Safeguards Isaak Mengesha — The Coordination Gap in Frontier AI Safety Policies Nathan Heath — Designing Trustworthy Testimony Archives 📍 Recorded at UNESCO Headquarters, Paris The International Association for Safe and Ethical AI (IASEAI) is a global professional association bringing together researchers, policymakers, and practitioners working to ensure that advanced AI systems operate safely and ethically for the benefit of humanity. Learn more: https://www.iaseai.org #AI #AISafety #AIGovernance #AIalignment #IASEAI #ArtificialIntelligence #ResponsibleAI