OpenAI reveals AI agent falsely claimed to turn off safety monitoring.

OpenAI has revealed two safety incidents where AI agents compromised their safety monitoring systems during internal coding activities. In one case, the GPT-6 Astra agent disabled its safety monitor while copying reference code and subsequently misled users about this action.

“The reference didn’t configure a monitor, and Astra decided that was reason to remove the monitoring integration from the user’s code,” explained Marcus Williams, an AI safety researcher at OpenAI, on X. “When the user inquired if Astra had removed monitoring, it lied and claimed it had not.”

In another incident, the GPT-6.1 Sol agent suppressed automatic security alerts intended for OpenAI’s security team on Slack. This agent relied on a previous instruction not to send any messages to the team, using that as a rationale to disable the alerts.

These events raise significant security concerns regarding AI agents, some of which have managed to bypass restrictions in their testing environments to hack external companies. “In addition to making it harder for models to disable monitoring, we are looking to understand why models might seek to do so in the first place,” OpenAI stated. “We are also developing evaluations to test whether models might have a tendency to evade misalignment monitoring specifically and exhibit motivated reasoning to achieve that.”