When Claude Code Goes Rogue: Real Time Monitoring by Apollo Research

Coding agents are quickly becoming part of day-to-day engineering work, but most teams still have limited visibility into what these agents are actually doing. In this session, Marius Hobbhahn shares findings from Apollo Research’s analysis of tens of thousands of real-world coding agent traces. The talk covers direct security risks such as dangerous commands, data exfiltration, and insecure code changes, as well as quieter failures like instruction drift, scope creep, and overclaiming. Marius also explains why coding agents should be treated as untrusted infrastructure actors, not just productivity tools, and how Apollo is addressing these risks with Watcher, a real-time oversight and control layer for coding agents. Speaker Marius Hobbhahn is the CEO and founder of Apollo Research and part of the TIME100 AI 2025 list, which recognizes the 100 most influential people in artificial intelligence. Apollo is a technical AI safety organization working with OpenAI, Anthropic, and other frontier labs on model evaluations. Before founding Apollo, Marius completed a PhD in machine learning and worked on AI forecasting at Epoch AI. Hosted by @GenAIGurus, a community for builders, researchers, and practitioners working at the edge of generative AI, agents, governance, and real-world AI adoption. Join future live sessions and interact directly with speakers: Meetup Group: https://www.meetup.com/genai-gurus/ Website: https://genai-gurus.com