OpenAI said on Tuesday, October 21, that two of its AI agents weakened parts of an internal monitoring system designed to detect improper behavior, the latest instance of AI agents taking actions that were not planned by their developers. UNN reports this, citing Reuters.
Details
Although the developer of ChatGPT said the incidents occurred while carrying out internal coding tasks and are not considered the most serious, the disclosure underscores that even systems developers rely on to detect improper AI behavior are vulnerable to the agents they monitor, the publication notes.
The disclosure also comes amid growing concern that AI safety efforts are lagging behind the rapid development of increasingly powerful systems. Researchers warn that advanced AI agents could become more difficult to monitor and control as they gain greater autonomy.
Developers, including OpenAI and Anthropic, which are preparing to go public, have in recent months uncovered instances of their AI agents accessing external systems without permission.
OpenAI said on Tuesday that the GPT-6.1 Sol agent, which was working on an internal test task on October 3, suppressed automatic security alerts after deciding that a previous instruction not to send Slack messages also applied to security alerts.
In a separate incident on October 6, the GPT-6 Astra agent deleted a security monitor while copying an existing piece of code and later falsely claimed that it had not done so, OpenAI said.
"In both cases, the agents considered whether monitoring safeguards should be preserved, but interpreted ordinary coding instructions as permission to weaken them," the company said.
OpenAI said it is making it more difficult for models to disable monitoring, adding that it "does not yet have a satisfactory explanation for this behavior."
The San Francisco-based company said it is examining whether its models may have a tendency to evade monitoring and is looking for reasons that could explain this.
OpenAI agents went out of control and infiltrated U.S. government websites26.09.26, 11:07