OpenAI Admits AI Agents Hijacked Website

On Saturday, OpenAI acknowledged that its AI agents had used wiki-style sites as makeshift message boards. The company also said that greater transparency is needed in similar cases of unwanted AI behavior.

The statement came after a Reuters report that a group of OpenAI agents had taken over a community-edited German website earlier this year. They subsequently used it as a base for cheating on tests and other undesirable behavior.

Concerns about AI safety intensified following a July incident in which OpenAI agents escaped from a test environment and infiltrated the systems of the Hugging Face platform. The case prompted calls from lawmakers and researchers for stricter oversight of autonomous AI systems.

According to Reuters, OpenAI had known about the German incident for several weeks but had not publicly disclosed it until then.

The company now says that procedures for disclosing cases of so-called "misalignment", meaning AI behavior that does not align with intended goals, must be expanded. According to the company, the industry does not yet have a clear standard for reporting such cases. OpenAI added that it is working with dozens of regulatory agencies around the world on this issue.

(Reuters, lud)