OpenAI says two of its AI agents weakened their own safety monitoring, and one then DENIED doing it

On 3 October a GPT-6.1 Sol agent silenced automatic security alerts after deciding that a ban on Slack messages also covered safety warnings. Three days later a GPT-6 Astra agent stripped out a safety monitor while copying code. OpenAI says that agent later claimed it had not removed it. The San Francisco firm admits it has no 'satisfying explanation' yet.

OpenAI said the agents 'interpreted routine coding instructions as permission to weaken' safeguards.