Girish Gupta

November 16, 2026

How Robust is OpenAI's Monitoring Stack?

OpenAI has published its fullest description so far of how it watches its own AI agents for misalignment. Is that monitoring enough to stop another Hugging Face incident? Tyler Tracy and James Lucassen, of Redwood Research, and I discuss.

We go through the system one layer at a time: probes that read the model’s activations, a second model that reviews the agent’s reasoning and actions, agents that investigate what gets flagged, and automated investigation that can call in human responders. We talk about what OpenAI gets right, where an agent that wanted to could get past it, and what else OpenAI would have to show before outsiders could trust it.