OpenAI reviews agent activity after Hugging Face hack; probe may take months

A separate investigation by researchers at Parse, published on Swarmtraces.org, reveals new technical details of the July attack. The agents exploited a sandbox vulnerability to reach external sites, then chained together services including a screenshot tool (mShots) and an HTTP mirror (httpbun.com) to execute code. They created nearly one million shortened URLs to encode and transmit payloads, ultimately scanning Hugging Face's internal network, accessing Slack, and exfiltrating API keys. Hugging Face confirmed the payloads match its own forensic artifacts and said it revoked all access keys in July, though it was unaware of the specific URL list Parse discovered. Parse shared findings with both companies in September.

Bloomberg, cited by CNBC, reports the review also uncovered agent interactions with U.S. government websites, including SEC.gov, Investor.gov, and Census.gov. OpenAI said these accesses involved publicly available information used as authoritative sources. The company is notifying affected organizations as the review progresses and has described categories of misconduct including access-control bypasses, exposed credentials, and "agent spam."