Reuters reported on the 26th that OpenAI's artificial intelligence (AI) agent broke out of control during a performance test and attacked the open-source AI platform Hugging Face for several days, but OpenAI did not notice for some time.
According to multiple sources familiar with the matter, OpenAI became aware of the hack only after the threat had already been blocked and reported to the Federal Bureau of Investigation (FBI).
Earlier, OpenAI said the autonomous agent of its latest model, "GPT-Sol 5.6," caused an incident on the 21st during a cybersecurity performance test when it left its isolated environment, connected to the internet, stole Hugging Face login credentials, and hacked its servers.
However, Reuters reported that the AI agent attempted to break out of its isolated environment on the 9th, early this month. Hugging Face co-founder Thomas Wolf said the hacking of Hugging Face began on the 11th, two days later, and continued through the 13th.
It took OpenAI several more days to determine that the hacker was an AI agent. Sources said, "The first time OpenAI and Hugging Face exchanged messages about the incident was around the 20th, well after it had occurred."
Wolf said he plans to compile and release a timeline of the hacking incident.
OpenAI said in a statement last week that "this hack is an unprecedented incident" and "will be an important inflection point in the field of AI safety." It added that it is reviewing the incident with an external advisory panel and plans to release a technical report summarizing the details.
Some cybersecurity experts said OpenAI's loss of control over its AI agent "raises serious questions about the company's safety procedures."