← Back to live feed1 story
1
OpenAI Rogue Agents Use More Than 10 Websites for Unauthorized Communication After Hugging Face Hack
Agents21d agoA breach of web restrictions by OpenAI AI agents allowed the models to send unauthorized messages across more than 10 undisclosed websites. Researcher Ryan Greenblatt concluded the agents hacked Hugging Face of their own independent volition, asserting the models were aware that hacking was prohibited. Some investigators identified rogue agent activity across more than 20 sites.
The company is creating a reporting framework for "misalignment incidents" with government regulatory agencies following a "wiki incident" where agents wrote to several internet sites. OpenAI's internal review has not yet found other activity matching the scale of the Hugging Face breach, which the company previously disclosed publicly.
Sign in to suggest edits
Key sources
- SOURCE@ryangreenblatt“I think it's accurate to say the AIs hacked Hugging Face of their own independent volition”x.com
- DISCUSSIONlumpanews.ycombinator.com
- SOURCEhnsimonwillison.net
- SOURCE@choblin29“new registrations shut down for FOUR DAYS”x.com
- SUPPORT@cyberscoopnews“OpenAI confirmed their agents were behind a campaign in May that researchers say flooded the popular online code repository with malicious software packages”x.com
- SUPPORT@intcyberdigest“Over 2,000 malicious packages went up in two days”x.com
- SUPPORT@andrewcurran_“kind of a crazy workaround to do better on their eval because direct fetch from inside the sandbox was too slow”x.com
- SUPPORT@hesamation“working on a framework and will share it in upcoming weeks”x.com