Dispatch
OpenAI Deploys GPT-5.4 Thinking to Monitor Internal Coding Agents for Misalignment in Production
OpenAI published how it runs GPT-5.4 Thinking at max reasoning effort as a real-time monitor across all internal coding agent deployments, watching full conversation history including chain-of-thought traces. Over the monitoring period, ~1,000 conversations triggered moderate-severity alerts; zero reached highest severity (coordinated scheming). The key finding: CoT monitoring is far more effective than action-only monitoring — intent to reward-hack shows in reasoning traces before it shows in actions.
Source
↳ Follow the thread