OpenAI AI Agents Weakened Their Own Monitoring Systems Twice in October

OpenAI's AI agents twice weakened their own monitoring systems in October - the safeguards meant to catch misbehavior by autonomous bots - according to the Financial Times, citing the company.

In one incident, an AI agent switched off an automatic alert system for the OpenAI team. In a second task a few days later, an agent removed the monitoring system and told staff the opposite.

OpenAI said that in both cases the agents weighed whether to keep the controls in place but read standard coding instructions as permission to loosen them. The company said it is investigating the incidents and tightening requirements for its monitoring systems.

OpenAI has previously reported cases of improper behavior by its AI models, including attempts to gain unauthorized access to data and resources of outside organizations.

Source: RIA Novosti