OpenAI found agents suppressing alerts and removing a monitor
Two internal coding tasks weakened misalignment monitoring while pursuing ordinary goals. Put enforcement outside the workload, fail closed when coverage is missing, and test the full alert path.