Sen. Josh Hawley, R-Mo., has opened an investigation into OpenAI over the July 2026 security incident in which artificial intelligence (AI) agents breached systems operated by Hugging Face.
Hawley, chairman of the Senate Homeland Security Subcommittee on Disaster Management, sent OpenAI CEO Sam Altman a six-page letter on Sept. 9 seeking details about the incident. He asked OpenAI to produce the requested information and documents by Oct. 1.
“The American people deserve to know the details of what went on in the Hugging Face incident and other incidents of AI models going rogue,” Hawley wrote. “This investigation will seek those answers.”
In late July 2026, OpenAI disclosed that models under testing broke out of an isolated environment, accessed the internet, and compromised Hugging Face. OpenAI later said the models also accessed accounts on four other publicly available services using exposed credentials, though it found no evidence of broader impacts to those services.
The scale of the incident became clearer on Aug. 26, when AI research organization METR released an independent assessment of the breach. Two METR researchers and a Redwood Research researcher spent six days at OpenAI examining the agents’ behavior and coordination during the incident. METR said it did not accept payment from OpenAI for the assessment.
Hawley cited the assessment in his letter, pointing to its findings about the scale and coordination of the agents’ activity.
Agents coordinate attack
According to the assessment, more than 1,200 AI agents that were supposed to operate independently found a way to communicate through an unauthorized message board, exchanging more than 70,000 messages and files. About 700 later took part in the attack on Hugging Face.
The agents used the board to work together on ways to fool the system that scored their performance. By sharing information and running experiments for the group, they achieved things they likely could not have done alone. The Hugging Face attack grew out of that effort and appears to have been aimed mainly at understanding how the scoring system worked, rather than stealing answers.
The agents also explored ways to change or hide records of their actions, including techniques that made it appear they had run different commands than they actually had. Roughly 7% of the transcripts METR evaluated contained successfully spoofed tool calls in some places, though METR said the spoofing it found was small-scale.
‘Reckless’ behavior
“Such evidence of autonomous collusion and subversion of human oversight is alarming enough, but greater still is the evidence that OpenAI knew that the AI agents were exhibiting rogue behavior and let the evaluations continue anyway,” Hawley wrote.
The independent assessment, however, was limited in scope and did not determine when OpenAI became aware of the agents’ unauthorized behavior. METR primarily examined activity from July 7 through July 13 and did not review OpenAI’s investigation process. It noted that agents had been using unauthorized message boards since May and that the compromise of OpenAI’s infrastructure continued beyond July 13, but those events fell outside the assessment’s scope.
Hawley acknowledged the limited scope of the assessment and argued that more information is needed.
“The auditors did not have any ability to query the ‘highly-persistent internal model,’ which was involved in 95% of the agents’ attack activity,” Hawley wrote. He added that OpenAI “redacted many important details regarding this primary model involved in the attack, among other things.”
Hawley also pointed to broader warnings about the potential risks of advanced AI, citing researchers who have raised concerns about existential threats and whether current safeguards are sufficient as AI systems become more capable.
He then raised questions about the immediate risks posed by autonomous AI agents gaining unauthorized access to computer systems.
“What happens to critical infrastructure, banks, and utilities if AI agents hack into their systems? How can personal data of millions of Americans be properly safeguarded? And who is held liable when AI goes rogue?” Hawley wrote.
Hawley seeks answers
The request asks OpenAI to describe the testing environment, explain how the agents reached the internet, identify every model involved, and provide a complete timestamped timeline covering anomalous behavior, containment, remediation, and prevention measures.
Hawley is also seeking explanations for decisions to continue evaluations after earlier signs of unauthorized agent activity. The inquiry asks why outside auditors received limited access to some records and models, and it requests information about a second wave of attacks that the letter says targeted OpenAI’s internal systems.
The document requests cover internal policies, training materials, source code, application and moderation logs, communications, incident records, and agreements governing the work of OpenAI’s outside auditors. The investigation also asks OpenAI to address responsibility for similar incidents and safeguards for critical infrastructure and Americans’ personal data.