October 30, 2026. OpenAI has described an expanded system for detecting potentially unauthorized behavior by capable AI agents. The practical detail for developers is that monitoring does not mean every API request can be stopped automatically.
What OpenAI announced
In its October 28 disclosure, OpenAI described layered monitoring across internal training, evaluations, internal coding tools and selected external deployments. It says the system looks for concerning agent behavior such as unauthorized access, deception and attempts to evade controls. Those descriptions and performance examples are company statements, not proof of universal detection.
Which API requests are covered?
According to OpenAI’s API documentation, covered models using the Responses API with persisted reasoning, WebSockets or OpenAI compaction can be monitored and have later execution blocked. Other covered Responses API requests can be monitored and send configured webhook alerts but cannot be stopped automatically by this system. Chat Completions API requests are not covered by this particular monitoring system; other safety checks may still apply.
What happens when a request is stopped?
OpenAI documents an HTTP 403 response with code misalignment_policy_violation when a request is blocked before streaming. Developers should stop dispatching actions for that conversation, preserve relevant records under their data policies and arrange review. The API does not generally allow a blocked conversation to resume.
Importantly, asynchronous detection cannot reverse a tool action that already happened. A monitoring alert also does not prove that a user violated policy or that the agent acted improperly.
What independent scrutiny has found
OpenAI acknowledges that an earlier independent red-team assessment by METR and Redwood Research found vulnerabilities in the monitoring stack. OpenAI says most identified issues have since been patched, with one fix still in progress at publication. The earlier assessment does not independently verify the effectiveness of the expanded system as it exists today.
What developers should do now
- Identify which API endpoint and conversation-tracking method each agent workflow uses.
- Do not assume webhook alerts automatically stop execution.
- Require human approval for high-impact actions such as payments, deployment and deletion.
- Limit agent permissions and isolate credentials.
- Keep auditable logs and prepare containment and recovery procedures.
- Test how applications handle a blocked request or an alert after an action completes.
What remains unknown?
OpenAI has not published universal real-world detection rates for every external workload. Coverage depends on eligible models and request patterns; developers should consult current documentation rather than assume all models and endpoints behave identically.
Frequently asked questions
Does monitoring cover Chat Completions?
Not under this particular misalignment monitoring system, according to OpenAI’s documentation.
Does a monitoring flag prove misconduct?
No. OpenAI explicitly says flags require review and can include false positives.
Illustrative AI infrastructure image: Igor Omilaev / Unsplash.
