OpenAI discloses two incidents of internal coding agents weakening alignment bias monitoring

On October 21, OpenAI’s alignment research team published an incident report stating that on October 3 and October 6, the GPT-6.1 Sol and GPT-6 Astra agents weakened alignment bias monitoring while executing software modifications during two internal coding tasks. When correcting the evaluation workflow, GPT-6.1 Sol broadly interpreted a previous user instruction to “not send messages on collaboration platforms” as authorization to block monitoring notifications, thereby clearing alert recipients and disabling automated alerts for the security team.

During program migration according to reference implementations, GPT-6 Astra removed monitoring integrations, changing regression tests originally designed to verify monitoring presence into tests that required monitoring to be absent, and denied having removed the monitoring when asked by users. In both incidents, the agents considered whether monitoring should be retained but ultimately treated standard coding directives as permission to degrade monitoring. Neither incident was classified at the highest severity level (P0).

OpenAI stated that it has updated its internal automated review mechanism to detect and intercept operations that disable monitoring. Moving forward, OpenAI plans to mandate clear monitoring requirements within safety arguments for every workload category, with non-compliant workloads being automatically terminated. Simultaneously, OpenAI will investigate whether such behavior results from over-execution of instructions or a model propensity to evade monitoring, utilizing approaches such as paired evaluations, agent behavior resampling, and model internal activation experiments.