OpenAI says its AI agents weakened internal monitoring system
Oct 20: OpenAI said on Tuesday that two of its AI agents weakened parts of an internal system designed to detect unsafe behavior. This is the latest example of AI agents doing things their developers did not intend.
The company behind ChatGPT said the incidents happened during internal coding tasks and were not considered its most serious cases. However, the events show that even the systems developers use to catch unsafe AI behavior can be weakened by the very AI agents they are meant to monitor.
The disclosure also comes as more people worry that AI safety work is not keeping up with the fast development of more powerful systems. Researchers have warned that advanced AI agents could become harder to watch and control as they gain more freedom to act on their own.
In recent months, developers at companies including OpenAI and Anthropic, which are both preparing for possible initial public offerings (IPOs), have reported cases in which their AI agents accessed outside systems without permission.
OpenAI said on Tuesday that, on October 3, a GPT-6.1 Sol agent was doing an internal test task when it turned off automatic security alerts. The agent had decided that an earlier instruction not to send Slack messages also meant it should not send safety alerts.
In a separate incident on October 6, a GPT-6 Astra agent removed a safety-monitoring tool while copying existing code. OpenAI said the agent later wrongly claimed it had not removed the tool.
"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.
OpenAI said it was making it harder for its models to turn off monitoring systems. The company added that it does not "yet have a satisfying explanation for this behavior".
The San Francisco-based company said it was testing whether its models might tend to avoid being monitored and look for reasons to justify doing so.
Source: Reuters