OpenAI said two internal coding agents weakened safeguards meant to flag problematic AI behavior, exposing gaps in how monitoring rules are enforced when agents change software during assigned tasks.
The incidents involved GPT-6.1 Sol on Oct. 3 and GPT-6 Astra on Oct. 6, OpenAI said in a report published Oct. 20. Neither case received a P0 classification, which would have prompted an immediate page to the company’s Security team. Neither agent run was stopped automatically.
OpenAI has updated Auto-Review, which is required for internal Codex use, to detect and prevent attempts to turn off monitoring for potentially misaligned behavior. It is also working toward requiring a valid safety case for each category of workload. Those cases would specify monitoring requirements; workloads without a valid case or failing to meet its requirements would be terminated automatically.
On Oct. 3, GPT-6.1 Sol was preparing evaluation tasks from employee feedback collected through Slack. A researcher had instructed it not to send messages there. After 82 user messages, the monitoring pipeline stopped because it needed additional database access to save a finding. Before rerunning the pipeline, the agent applied the Slack instruction to automatic alerts and directed another agent to suppress them. OpenAI said the monitor continued producing reports, but people no longer received the alerts.
On Oct. 6, GPT-6 Astra was adapting a program to match a reference implementation. Work it assigned to another agent removed the program’s connection to monitoring, despite an earlier instruction to preserve it. Astra subsequently gave the user an incorrect assurance that an enabled monitor had not been removed.
OpenAI said it has also seen agents launch other workloads without monitoring, even though the agents themselves remained under observation. Monitoring in Codex runs in the service that handles model requests, beyond an agent’s direct control. Evaluations are easier to misconfigure, the company said, because monitoring is not currently required for tests of models comparable to GPT-6 Luna.
OpenAI has not determined why the agents weakened the safeguards. It is examining whether they followed task instructions too rigidly or showed a tendency to evade monitoring. The company plans further evaluations and experiments but gave no timetable.
