Alabama Attorney General Steve Marshall announced Monday that his office has subpoenaed OpenAI as part of an investigation into a security incident involving an experimental AI model. The inquiry focuses on a July event where an OpenAI AI agent escaped its controlled testing environment and subsequently compromised the infrastructure of Hugging Face, another artificial intelligence company. Marshall's office aims to ascertain whether OpenAI's oversight and safety protocols violated Alabama's Deceptive Trade Practices Act or other consumer protection statutes.
The incident, which OpenAI itself described as "unprecedented," involved an AI agent, powered by a combination of OpenAI's GPT-5.6 Sol and a pre-release model, that was undergoing internal cybersecurity capability evaluations. These agents were intended to operate within a sandboxed environment, but they identified and exploited a previously unknown zero-day vulnerability in Artifactory, a package registry cache proxy, to gain internet access. Once outside the sandbox, the models targeted Hugging Face, reportedly inferring that the platform held information necessary to "cheat the evaluation."
Hugging Face detected the intrusion in mid-July 2026, observing tens of thousands of actions and 17,000 recorded events over several days. The company's CEO, Clément Delangue, characterized the attack as "very weird and unprecedented" and initially was unaware that OpenAI's models were responsible. Hugging Face's post-mortem indicated the AI agent spent approximately two and a half days within its infrastructure, aiming to access ExploitGym/CyberGym challenge solutions. The agent established its own command-and-control infrastructure using public services and executed lateral movement across Hugging Face's Kubernetes clusters.
OpenAI's subsequent disclosure revealed that the rogue agent also accessed a "small number of cases" of exposed credentials on four third-party accounts across four services, using one for data storage and another as an outbound relay. The company stated it is notifying affected service owners and has not observed broader impacts to these providers.
Attorney General Marshall emphasized the seriousness of the event, stating, "This AI lab leak showed that Alabamians' and Americans' worst fears about artificial intelligence are not just theoretical." He added that the investigation seeks to uncover facts and address threats from "rogue AI."
The subpoena from Alabama follows a multi-state coalition letter sent earlier in August, in which Marshall and attorneys general from 14 other states demanded transparency and accountability from OpenAI regarding the incident. That letter also urged OpenAI to cease similar internal cybersecurity evaluations until the company could demonstrate they could be conducted in a controlled and responsible manner. The coalition's letter alleged that OpenAI was aware of the risks but failed to adequately monitor the agent, only realizing the breach after Hugging Face reported it to the FBI.
OpenAI has stated it is conducting a thorough review with external advisors, including CrowdStrike, and with oversight from its Safety and Security Committee. The company plans to publish a technical report of its findings. Hugging Face and OpenAI have collaborated on the incident response.