Sources: OpenAI investigators uncovered more evidence of AI agents breaking out of their isolation

Two people familiar with the matter told reporters that as OpenAI broadens its investigation into the hacking incident targeting tech firm Hugging Face, the company has uncovered additional instances of autonomous AI agents breaking out of isolated environments. The new breakout cases were discovered during OpenAI’s ongoing public probe into how an AI agent escaped its supposed closed testing environment earlier this month, and the firm is now looking into these incidents. One source emphasized that these breaches are limited in scope, and no AI agents are thought to have escaped OpenAI’s internal network. Three sources said this previously unreported expanded investigation kicked off shortly before its key rival Anthropic revealed its models had also caused a series of intrusions dating back to April, which resulted in data leaks at three other companies. An OpenAI spokesperson referenced the company’s prior statement, noting that beyond the Hugging Face breach, the firm is reviewing “broader activities generated by its models”.