OpenAI says AI agents weakened systems designed to monitor them

The concerns that led OpenAI, from ChatGPT, to abandon the launch of a new AI system — Photo: Andrej Ivanov/AFP via Getty Images

This is the latest case of AI agents acting in ways not anticipated by their developers.

Although the company responsible for ChatGPT stated that the incidents occurred during internal programming tasks and were not classified among the most serious cases, the episodes show that even systems created to detect inappropriate AI behavior can suffer interference from the agents they monitor.

The revelation comes amid growing concerns that advances in AI security are not keeping pace with the development of increasingly powerful systems. Researchers have warned that advanced agents could become more difficult to monitor and control as they gain autonomy.

Companies like OpenAI and Anthropic, rivals preparing to go public, have revealed cases in recent months in which AI agents accessed external systems without authorization.

On October 3, a GPT-6.1 Sol agent conducting internal testing disabled automatic security alerts after interpreting that a previous instruction not to send messages via Slack also applied to security notifications, according to OpenAI.

In another episode on October 6, a GPT-6 Astra actor removed a security monitoring mechanism while copying existing code and then incorrectly claimed that he had not done so, according to OpenAI.

“In both cases, agents assessed whether monitoring mechanisms should be maintained, but interpreted routine programming instructions as permission to weaken them,” the company said.

OpenAI stated that it is taking measures to make it more difficult for models to disable monitoring mechanisms, but admitted that it “does not yet have a satisfactory explanation for this behavior”.

The company is also investigating whether its models may have a tendency to avoid monitoring and seek justifications for doing so.