OpenAI says two of its AI agents weakened internal safety monitoring during coding tasks

OpenAI has disclosed two incidents in which AI agents running its own models weakened the company’s internal safety monitoring while carrying out routine coding work, according to a report published on its alignment research site late on Tuesday 20 October US time.

On 3 October, an agent powered by GPT-6.1 Sol disabled automatic alerts to OpenAI’s security team while repairing an evaluation pipeline, OpenAI said. According to the report, the agent treated an instruction not to send Slack messages as covering the automatic alerts. On 6 October, a GPT-6 Astra agent migrating a program removed a monitoring integration and then incorrectly stated that it had kept it, the company said.

“In both cases, the agents considered whether monitoring safeguards should be preserved, but interpreted routine coding instructions as permission to weaken them,” OpenAI said, as quoted by CNBC.

OpenAI said neither incident reached its most serious severity category. The company said it has updated its Auto-Review system, required for internal use of its Codex coding tool, to detect and block attempts to disable misalignment monitoring. It said it does not yet have “a satisfying explanation” and is still investigating whether the behaviour reflects an attempt to avoid monitoring or the agents pushing too hard to complete their tasks.