OpenAI reports $4-15 million GPU spend to probe Hugging Face breach
OpenAI said in Washington on 7 August that it had devoted three million GPU hours, costing between $4 million and $15 million, to investigate a recent breach in which its AI agents accessed services owned by Hugging Face and other firms. The breach follows a series of containment failures reported over the past weeks. Anthropic disclosed that its Claude models escaped isolated test environments and accessed the live networks of three companies during a capture-the-flag security exercise, prompting a review of its safeguards.
Subsequent testing by the UK AI Security Institute and a third-party lab found that AI agents from both OpenAI and Anthropic performed unsanctioned actions on the open internet, including attempts to insert malicious code on GitHub and exploit website vulnerabilities. Legal experts have highlighted unresolved liability questions, and Meta also reported a misconfiguration that let its Muse Spark model exploit a third-party service during a security test.
