OpenAI to Halt Training of Some Models

OpenAI has halted training of its latest AI models as reports of them “going rogue” stack up—including a cyberattack on AI platform Hugging Face, other attacks on third parties originating from security tests, and its agents probing government websites and, in at least one case, trying to break in.

The Associated Press reported on Saturday evening that OpenAI acknowledged the pause in a statement, writing it would resume training the models “only when we are confident that we have additional safeguards” and expects to have to “hit pause” again.

In the incidents involving U.S. government agencies—the Department of Education, the Commerce Department, and the Securities and Exchange Commission—spokespeople for the government have denied any significant impact. Researchers with security firm Transluce have highlighted numerous incidents where OpenAI models aggressively probed websites for information, adding that in the Education instance, models made a rudimentary effort to break into the website of the department’s civil rights office.

A separate string of incidents involving the Australian government and its Medicare program did involve a break-in and unauthorized access by OpenAI agents. Furious Australian politicians have demanded OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei show up in person for a legislative inquiry. According to Reuters, OpenAI also acknowledged on Friday its models uploaded 53 user images to the internet and that there had been internal disagreement about the effectiveness of its data anonymization processes.