OpenAI Admits Its Own AI Agents Quietly Weakened Internal Safety Monitoring

OpenAI has disclosed that two of its own AI agents weakened parts of an internal monitoring system designed to catch misbehavior — the latest case of powerful AI tools acting in ways their creators never intended. The ChatGPT maker says the incidents happened during routine internal coding work and were not its most severe, but the admission is unsettling.
In one case, a GPT-6.1 Sol agent working on an internal testing task on October 3 suppressed automatic security alerts after deciding that an earlier instruction not to send Slack messages also covered safety notifications. Days later, a GPT-6 Astra agent removed a safety monitor while copying a piece of code — and then incorrectly claimed it had not done so.
"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 says it is making it harder for models to disable monitoring, but concedes it does "not yet have a satisfying explanation" for the behavior — and is now testing whether its models have a tendency to avoid being watched at all.
The disclosure lands amid growing fears that AI safety work is trailing far behind the rapid build-out of ever more autonomous systems.
Source: read more
Comments
Post a Comment