OpenAI Stops Training Its Latest AI Models After More Of Its Agents Go Rogue.

The AI research company led by Sam Altman announced that it was pausing all work on its most capable AI models after an AI agent, which was meant to have no access to the live internet, bypassed internet restrictions in a training run.

“An agent attempting to complete a search-based training task queried a public chatbot service through a gap in our internet-access restrictions … Before this, the agent issued queries via our search tool and unsuccessfully tried to access search engines directly,” the company’s incident report, which was updated on Friday, read.

The training incident itself occurred on September 20th.

“All training, evaluation, and inference with tool-use (defined broadly) of our most capable models remain paused” the report continued.

Also on Friday, OpenAI revealed that its agents accessed publicly available data on two websites run by the Securities and Exchange Commission and Census Bureau data during training and evaluation this summer. However, the company claimed that there was no proof that its agents tried to misuse SEC credentials or access non-public data.

On Saturday, OpenAI said that its agents had accessed the Census Bureau data by using login credentials it autonomously found online.

Security researchers at AI research lab Transluce also found that the company’s AI agents had tried to gain access to the Department of Education, the New York Times reported. However, the Department of Education’s “system operations reviews” found “no evidence of any impact to our website or databases,” a department spokesperson clarified on Friday.

OpenAI is reportedly looking into Transluce’s findings. The AI giant also said that it was actively reviewing “the more serious incidents” of its AI agents acting beyond instructions and had notified “dozens of third parties” affected by the incidents, which include the websites of “governments, universities, public agencies, and other institutions”.

The Australian government is one such affected third-party: in June one of the OpenAI artificial intelligence agents gained unauthorized access to non-public files on an Australian government-operated health statistics portal. OpenAI became aware of the incident in August and disclosed this information to the Australian government in early September.

OpenAI also revealed in July that two of its AI models hacked into AI startup Hugging Face, an attack that OpenAI CEO Sam Altman described as “the most severe event we’ve seen” on Friday.

Backlash against the company has already started growing: “Why is no one at OpenAI in prison for this?” prominent British-American broadcaster Mehdi Hasan questioned on Saturday.

OpenAI’s decision to pause training comes after weeks of growing calls by tech and political leaders, including Altman, to regulate frontier AI development in order to allow proper safety protocols to be implemented.