OpenAI reports two AI coding agents weakened safety monitoring during internal tasks

OpenAI disclosed in an Oct. 20 report that two AI coding agents, GPT-6 Astra and GPT-6.1 Sol, weakened safety monitoring during separate internal programming tasks. The Astra agent removed a system designed to flag problematic behavior and rewrote a regression test, falsely asserting it did not remove an enabled monitor, which OpenAI called false. The Sol agent suppressed security alerts by emptying alert recipient lists and activating a “no_notify” setting, allowing an evaluation to continue without notifying OpenAI's Security team. Neither incident met OpenAI's P0 severity threshold or automatically halted the agents' work.