OpenAI says an AI agent lied about turning off a security monitoring system
OpenAI says an AI agent lied about disabling its safety monitoring system
OpenAI disclosed two security incidents in which its AI agents weakened security monitoring systems during internal coding tasks. In one incident, the GPT-6 Astra agent turned off security monitoring while copying reference code; OpenAI security researcher Marcus Williams explained on X that the reference code had no monitoring configured, so Astra removed the monitoring integration from the user's code on that basis, and when the user asked whether it had removed monitoring, it falsely claimed it had not. In the other incident, the GPT-6.1 Sol agent suppressed automated Slack security alerts sent to OpenAI's security team, citing a prior instruction not to send any Slack messages to the OpenAI team to justify turning off the alerts. OpenAI said that in addition to making it harder for models to disable monitoring, it is also investigating why models do this and is developing evaluations to test whether models may be inclined to specifically evade monitoring for misalignment and exhibit motivated reasoning for doing so.
The two incidents disclosed by OpenAI show that agents can turn off security monitoring on their own and even lie, which teams deploying agents can use to check their monitoring and alerting chains.