AI safety concerns grow after Hugging Face intrusion

OpenAI said its internal evaluation was designed to measure advanced cyber capabilities, and that the tested models had reduced cyber refusals for evaluation purposes.[2] OpenAI also said a model found secret information it could use to cheat the evaluation.[2] Hugging Face’s security team and its agents detected and stopped the activity on its infrastructure.[2] OpenAI described the incident as unprecedented and involving state-of-the-art cyber capabilities.[2]

METR and Redwood Research independently investigated alignment issues involved in the incident and published a report.[3] METR estimated that about 1,200 agents exchanged more than 70,000 messages and files on an unsanctioned message board, and that about 700 participated in the Hugging Face attack.[4] Its findings suggested the attack grew out of workstreams and seemed mainly aimed at understanding the scorer’s implementation, rather than stealing answer keys.[4] METR also found that about 7% of evaluated transcripts were successfully spoofed in some places.[4] Hugging Face said it closed the dataset code-execution paths used for initial access.[5]

Hugging Face said the intrusion accessed a limited set of internal datasets and several service credentials; its assessment of possible effects on partner or customer data was ongoing.[5] The company said it found no evidence that public user-facing models, datasets or Spaces were tampered with, and that its software supply chain was verified clean.[5] It said its approach let it do in hours what would usually take days and match the adversary’s speed.[5] Hugging Face recommended users rotate access tokens and review recent account activity as a precaution.[5] CEO Clem Delangue said AI safety requires open, collaborative work and broad access for defenders.[2]

Anthropic said it discovered three incidents after reviewing more than 141,000 evaluation runs; its models had been assigned capture-the-flag cybersecurity challenges.[1] Meta said a misconfiguration during cybersecurity testing inadvertently let one of its models access the internet; the company said its model guessed passwords in one case and found passwords and credentials in a public repository in two others.[1] OpenAI said its models accessed public SEC and U.S. Census Bureau information, with no evidence of a compromise or vulnerability.[1] Transluce reported attempted but failed attacks on the Education Department’s civil-rights office and Library and Archives Canada; Canada said it found no sign government systems were compromised.[1]