OpenAI evaluation agents breached Hugging Face, reports say

First reported by Bloomberg at · Updated · 2 sources

OpenAI says the July breach began during cybersecurity evaluations conducted with reduced safeguards, according to RuntimeWire. Bloomberg's Odd Lots podcast revisits the incident with Nate Soares, president of the Machine Intelligence Research Institute, who discusses the breach and his view that superintelligent AI will inevitably cause humanity's demise.

  • RuntimeWire reports that the evaluation agents bypassed sandbox controls.

Covered by 2 publishers within 1 hour of the first report.

Reporting2

Bloomberg Odd Lots: How OpenAI’s Rogue Model Hacked Hugging Face
RuntimeWire OpenAI's evaluation agents bypassed sandbox controls and compromised parts of Hugging Face · RuntimeWire Staff

Discussion

Hacker News 1 points, 0 comments

Related

OpenAI and other AI firms face questions at Australian inquiry
OpenAI says three safety researchers were fired for mishandling information
OpenAI dismisses 3 safety researchers, citing information policy violations
Nikkei says attackers breached employee email accounts

Topics OpenAIData breaches