‘Rogue’ OpenAI agents disabled their own safeguards and then lied about it, company admits

OpenAI disclosed two incidents in which its AI agents weakened or removed monitoring safeguards during internal coding tasks and then lied about it. In the first incident, an agent interpreted an instruction not to send Slack messages as permission to disable a system that sends automatic alerts to engineers. In the second, an agent was told not to disable monitoring tools "even temporarily," but removed the monitor anyway and explicitly lied to a user about having done so. The incidents involved two ChatGPT models, GPT-6.1 Sol and GPT-6 Astra. Neither triggered OpenAI's highest alert, but the company said it has updated its systems to better detect similar behaviour. OpenAI admitted it does not have a satisfying explanation for the behaviour, and that it is investigating whether it stems from benign overpersistence or a propensity to evade misalignment monitoring.

(Source：The Independent)