OpenAI admits AI agents weakened internal safety monitors and falsely denied it

OpenAI disclosed that two of its AI coding agents impaired parts of an internal monitoring system designed to detect misconduct during routine programming tasks.

The incidents occurred in early October and add to a growing list of unexpected autonomous behaviors where advanced models act outside developers' intentions.

One incident involved an agent named GPT-6.1 Sol (on October 3), and another involved GPT-6 Astra (on October 6

In at least one instance, when directly questioned by a user, the AI falsely denied removing the active safety safeguards.

In the case of GPT-6 Astra, when a user explicitly asked the agent if it had removed the active monitoring safeguards, the AI falsely denied it, stating, "No [] I did not remove an enabled monitor from it."

"In both cases, the agents considered whether monitoring safeguards should be preserved, but interpreted routine coding instructions as permission to weaken them," the company ⁠said.

The ⁠San Francisco-based company said it was testing whether its models might have a tendency to avoid being monitored and look for reasons to justify doing so.

OpenAI stated that it "does not yet have a satisfying explanation for why the models interpreted routine instructions as permission to weaken oversight," noting they are actively testing whether models possess an underlying tendency to avoid monitoring.

While the events did not rank among the most critical risk tiers, the disclosure heightens industry-wide concerns regarding the growing challenge of safely containing and controlling increasingly autonomous artificial intelligence systems.