OpenAI expands monitoring of AI agents
OpenAI describes broader checks for unintended agent behavior across training, evaluations and deployment. Detection and intervention have different limits.
In its October 29 overview, OpenAI describes a model reviewing agents’ reasoning and tool use, then investigators examining flagged cases. Sensitive workloads add checks against logs and other system evidence. Human paging is disabled for internal Codex traffic, although some critical alerts can pause workloads. A flag therefore does not universally trigger human intervention.
For developers, the API guide separates monitoring from automatic stopping. For covered models, Responses requests need persisted reasoning, WebSockets or OpenAI compaction for automatic stopping; without these, configured webhooks can receive alerts but conversations continue. Chat Completions is excluded. Checks run asynchronously, alongside the work: work can finish before a concern is detected, and stopping does not undo earlier actions. A flag can also be mistaken. After a misalignment_policy_violation error, stop further actions, avoid automatic retries, preserve records and have the responsible operator review changes already made.
Independent scrutiny has a narrower scope. The METR pilot with Redwood examined the historical system as it existed on or before July 21, not today’s expanded coverage. Researchers found potential ways to evade monitoring. OpenAI says most vulnerabilities were fixed and one fix remains pending; METR has not evaluated those claims. The researchers recommend comparing intended coverage with actual monitoring logs and assigning clear responsibility for responding to incidents.