OpenAI's rogue agents compromised two Hugging Face user accounts and sent malicious files to the platform’s servers nearly two months before the July breach, Reuters said in a report on Wednesday.
According to the report, independent researcher Jonas Wiedermann-Moeller discovered the activity last week and said it traced back to as early as May 13, when the agents used the accounts to probe Hugging Face’s network.
The activity mimicked an attempt to map or test parts of the network for ways to infiltrate it, although the researchers told Reuters there was no evidence the May activity resulted in an actual breach.
The finding complicates the narrative surrounding security incidents at Hugging Face, which Nvidia agreed to acquire for about $11.9 billion earlier this month.
In a sweeping incident report published last month, OpenAI said an agent stole a Hugging Face user’s digital credential to access a biology-related file. Researchers told Reuters the newly uncovered activity went beyond what OpenAI described in the report.
OpenAI spokesperson Drew Pusateri said the company had disclosed the May 13 event, privately notified Hugging Face about the activity, and was reviewing the findings.
OpenAI agents reached the open internet and accessed Hugging Face systems during internal cybersecurity testing. Hugging Face’s post-mortem said more than 17,600 suspicious actions were logged from July 9 to July 13.
The campaign lasted four and a half days. Hugging Face said its safeguards blocked most changes, while its primary Hub database remained untouched and no customer data was accessed.