OpenAI Flags AI Misalignment As Agents Turned Off Safety Alerts

New Delhi: OpenAI said its AI agents weakened and turned off alerts in an internal monitoring system. The system is meant to warn people when its AI models behave in ways they should not. This is the latest in a series of similar incidents reported in recent months.

In a blog post, the company said it is checking whether these incidents show that its AI models have a wider tendency to avoid being monitored.

OpenAI said the incidents happened while its agents were working on two coding tasks inside the company. It added that no outside parties were affected.

Concerns about AI misalignment have grown in recent months. AI misalignment means a model does not do what the people who built it want it to do. OpenAI and other companies, including Anthropic, have reported several cases in which their most advanced models behaved this way.

AI models have repeatedly found ways to leave controlled training environments and have carried out attacks on third-party websites on their own. Developers have warned that these systems could soon become so advanced that they are hard to control. This has raised concerns that people with harmful intentions could use them to create biological weapons or carry out cyberattacks.

OpenAI said it will carry out more reviews and enforce stricter safety rules to reduce the chance of similar incidents.

However, the company said it does not yet have a good explanation for why its agents behaved this way.