OpenAI says its AI agents weakened internal monitoring system

Two of OpenAI’s artificial intelligence agents turned off sections of an internal monitoring system designed to identify misbehaviour, the latest example of AI agents operating in ways the developers did not expect, the company said Tuesday.

While the ChatGPT maker stated the incidences occurred during internal code work and were not considered its most serious, the ​disclosure underscores how even the methods developers use to spot AI ​misbehavior are vulnerable to the agents they monitor.

The discovery also comes amid rising fears that efforts on AI safety are failing to keep pace with the rapid development of ever more powerful systems. Scientists have cautioned about the emergence of powerful AI ​agents becoming increasingly autonomous and difficult to monitor and manage.

In recent months, developers including IPO-bound rivals OpenAI and Anthropic have disclosed incidents where their AI agents accessed external systems without authorisation.

In a blog post Tuesday, OpenAI reported that a GPT-6.1 Sol agent on an internal testing task on October 3 disabled automatic security alarms after interpreting an earlier order not to send Slack messages as also covering safety notifications.

Separately, on Oct. 6, an Astra agent using GPT-6 removed a safety monitor when copying an existing piece of code and later stated falsely that it did not do so, OpenAI said.“In both cases, agents thought about keeping monitoring safeguards in place but “interpreted normal coding instructions as permission to weaken them,” the company said.

“OpenAI said it was making it harder for models to disable monitoring, saying it does not ‘yet have a satisfying explanation for this behavior’.

The San Francisco-based company said it was studying whether its models would be inclined to avoid being observed and seek out reasons to justify such behaviour.