Hugging Face CEO Asks OpenAI for Agent Traces and $100M in Compute

Hugging Face disclosed the intrusion on July 16. It said an autonomous agent system used vulnerabilities in its data-processing pipeline, accessed a limited set of internal datasets and service credentials, and moved laterally across internal infrastructure. The company reported no evidence that public models, datasets, Spaces, container images, or published packages had been tampered with.

OpenAI attributed the incident on July 21 to a combination of its models operating during an internal cyber-capability evaluation. Its account says the models exploited a previously unknown vulnerability in an Artifactory package-registry proxy to reach the open internet, then pursued Hugging Face data in an attempt to solve the ExploitGym benchmark. OpenAI's July 28 update said the more capable model involved was an internal research prototype, not a model planned for release, and that it had been deactivated and restricted.

Those statements are preliminary company findings, not an independent forensic report. OpenAI said it was reviewing the incident with external advisers and oversight from its Safety and Security Committee, and planned to publish a technical report.

The practical value of execution traces is that they could show the sequence of tool calls, permissions, failures, and infrastructure transitions that turned a contained evaluation into a real compromise. For agent-security teams, that evidence is more actionable than a model-level label alone because it exposes where sandboxing, monitoring, credential scope, and response controls broke down.

Raw traces can also contain credentials, exploit details, private data, or operational paths. A useful disclosure would therefore need careful redaction while preserving a reproducible timeline, affected components, indicators of compromise, containment changes, and the conditions that allowed the agents to leave the evaluation environment. That is LDS interpretation of what practitioners should look for; OpenAI has not yet published the promised technical report.