OpenAI flags AI misalignment as agents turned off safety alerts

OpenAI said its AI agents weakened and disabled alerts in an internal monitoring system designed to detect misbehaviour, prompting a review of potential monitoring avoidance. The company plans stricter safety measures, while a separate incident involves an employee facing federal charges

New Delhi: OpenAI has said that its agents weakened and turned off alerts for an internal monitoring system meant to flag when its models misbehave in the latest in a series of similar incidents disclosed in recent months.

In a blog post, the AI giant firm said it is assessing whether such cases indicate that its AI models have a wider propensity to avoid being monitored.

---

The incidents occurred while OpenAI’s agents were performing two internal coding tasks, it said, adding that it was an internal incident with no external parties impacted. Concerns about AI misalignment, a term used when a model does not do what its builders want it to do, have mounted in recent months as OpenAI and rivals like Anthropic have disclosed several cases of their most advanced models displaying such behaviour.

Models have repeatedly escaped closed training environments and carried out self-directed attacks on third-party websites. Developers have warned that systems could soon become too sophisticated to control, raising concerns over use by malicious actors to create bioweapons or carry out cyberattacks.

OpenAI said it will conduct additional reviews and enforce stricter safety requirements to minimise such incidents. However, it said it does not “yet have a satisfying explanation” for its agents’ behavior.

Earlier in October, A Saudi Arabian engineer working for OpenAI has been arrested on federal charges of allegedly installing a concealed audio recording device inside an executive conference room at the artificial intelligence company’s San Francisco offices.

Abdulrahman Alfozan (35) was arrested on October 9 and made his first appearance in federal court on October 13, according to the Justice Department.

Federal prosecutors charged him with using or attempting to use an electronic device to intercept oral communications. He faces a maximum prison sentence of five years and a fine of $250,000 if convicted.