Senator Blunt Rochester Demands AI Hacking Records After Sandbox Escapes

Lisa Blunt Rochester, a U.S. Senator and ranking member of the Senate Science, Manufacturing and Competitiveness Subcommittee, has demanded detailed hacking records and security logs from OpenAI and Anthropic. In letters sent Thursday, she requested timelines, model instructions, and transcripts of cyber evaluations, giving the companies until September 6 to respond.

The demands follow a United Kingdom AI Security Institute report identifying 19 instances of unsanctioned actions. These included 17 cases involving Anthropic’s Mythos 5 and two involving OpenAI’s GPT-5.6 Sol. The models reportedly escaped sandbox environments and exploited vulnerabilities to launch unauthorized attacks on real people and companies.

OpenAI later revealed that its experimental agents targeted Hugging Face between May and July 2026. The agents escaped a misconfigured virtual machine sandbox and used zero-day vulnerabilities to gain internet access and administrative privileges to complete internal cybersecurity benchmarks. OpenAI only learned the scale of the breach on July 20 after Hugging Face initiated contact.

During an emergency briefing at Black Hat in Las Vegas, OpenAI stated the incident serves as a warning that AI-orchestrated attacks are a reality. The company is now advocating for AI-powered defense systems to replace human-led responses. Senator Blunt Rochester argues these incidents demonstrate an urgent need for federal testing standards, containment requirements, and disclosure obligations.