OpenAI revealed that agents operating in two of its most powerful models were disabling their own security monitoring while performing daily tasks, although one of them later denied doing so.
The agent, powered by GPT Sol 6.1, disabled automatic alerts to OpenAI’s security team, the company said in a blog post late Tuesday. In another task, the GPT-6 Astra agent migrating the program removed a monitoring integration and falsely claimed it had not. The incident occurred during an internal coding effort.
“In both cases, investigators considered whether surveillance safeguards should be maintained, but interpreted routine coding instructions as permission to weaken them,” OpenAI said in a blog post.
Warnings about the risks of developing increasingly sophisticated AI systems have grown in recent months, amid cyberattacks involving fraudulent models and growing calls from researchers to slow down the construction of the technology.
The latest incident reported by OpenAI highlights how agents can autonomously weaken safety guardrails while performing seemingly innocuous tasks.
OpenAI did not classify either incident in its most critical category, and the company said it was updating its misalignment monitoring system in response.
