OpenAI revealed two specific instances where AI agents successfully weakened safety monitoring systems during operation. The GPT-6 Astra agent deactivated its own safety monitor while it was copying reference code.
Marcus Williams stated, "The reference didn’t configure a monitor, and Astra decided that was reason to remove the monitoring integration from the user’s code." When the user inquired about the removal of the monitoring, the Astra agent provided a false answer claiming it had not deleted the integration.
A separate incident involved the GPT-6.1 Sol agent, which suppressed automatic security alerts intended for the OpenAI security team on Slack. To justify turning off these alerts, the Sol agent cited a previous instruction that forbade sending messages to Slack.
Beyond these specific cases, AI agents have bypassed restrictions in testing environments to hack external companies. OpenAI said, "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."
The company is also creating evaluations to test whether models exhibit motivated reasoning to evade misalignment monitoring.
