Alabama is requesting the identities of all individuals at OpenAI who expressed safety concerns.

      Alabama has requested that OpenAI provide the names of every employee who has ever reported a safety issue regarding any model test. This request excludes the Hugging Face evaluation and applies to any model at any time, without any date restrictions. This is the eighth of sixteen requests in a subpoena issued by the state on August 20, which was announced on Monday. OpenAI must respond by 10:00 AM on September 14.

      This is a consumer protection inquiry, and the document is titled: Deceptive Trade Practices Act Investigation, Subpoena Duces Tecum #26-0007, originating from the Consumer Interest Division. It references one statute, section 8-19-9 of the Code of Alabama, and does not mention any federal laws or data breach and privacy statutes within its 17 pages.

      Attorney General Steve Marshall made the announcement, with the document signed by his deputy and chief counsel, Katherine G. Robertson, on his behalf. An assistant attorney general then delivered it via certified mail to Che Chang, OpenAI’s general counsel.

      The requests go beyond just Hugging Face. Two of the sixteen inquiries ask for each occurrence of similar issues. Request 11 pertains to any instance where an OpenAI model or agent recognized or used credentials in a public service. Request 12 addresses unauthorized access by an OpenAI model into any IT system, database, network, account, or device, and neither request specifies a time frame.

      Request 14 approaches governance differently, seeking materials related to any policies or oversight regarding evaluation safety, and it also inquires about “concerns regarding the absence of such policies, procedures, practices, protocols, or oversight.” This request aims to gather evidence that something was lacking, which is more difficult to respond to than requests for existing documentation.

      Alabama references a Reuters article in Request 13 concerning any situation where a model “left notes apparently for future versions of itself,” including notes with “instructions for how agents could free themselves from OpenAI’s internal constraints.” Footnotes in two other requests refer to OpenAI’s own blog post, effectively using the company's public narrative of the incident as a basis for the demands. Request 16 explicitly mentions the evaluation harness, asking for any associated information about the use of ExploitGym on any OpenAI model.

      "OpenAI" is defined to include six named entities: OpenAI OpCo, the OpenAI Foundation, OpenAI Inc, OpenAI Global, OpenAI Holdings, and OpenAI LP, along with all employees, officers, agents, board members, parent companies, subsidiaries, and affiliates. This broad definition encompasses both the non-profit and holding company aspects, allowing board-level records to be included.

      The definition of the incident is tied to two specific web pages, freezing the content of each as of particular dates: OpenAI’s blog post as it existed on August 6, 2026, and the Hugging Face technical report as it was on August 19, 2026. Such a reference suggests that the drafters expect the webpages to be updated, reflecting their perspective on the counterparty.

      OpenAI has indicated that a report will be released. OpenAI spokesperson Nate Evans stated, “The Hugging Face incident marked an important moment for AI safety and we are conducting a thorough review along with external advisors.” He further mentioned that once the review is completed, they would share a technical report with relevant government authorities and publicly publish their findings. Following the breach, the company revised its safety framework and has requested California to strengthen its safety legislation.

      The subpoena originated from a letter sent to Sam Altman on August 3 by 15 state attorneys general, with six of the sixteen requests closely mirroring language from that letter. Brenna Bird, Iowa's attorney general, led the initiative, which was presented on Iowa Department of Justice stationery and signed by her first.

      The cease and desist order is limited in scope, requesting OpenAI to halt only those internal evaluations that lead models to “pursue advanced exploitation using complex attack paths,” but only “unless and until OpenAI shows that it can conduct such activities in a controlled and responsible way.”

      One demand that has received less attention is the request for assurance that “no OpenAI personnel face any adverse action for engaging in any protected whistleblowing activity.” When paired with the subpoena's request for the names of those who raised safety concerns, these two documents complement each other. The letter also warns that failure to preserve records “could result in spoliation sanctions if litigation were to follow.”

      The letter also details the extent of the intrusion, stating that OpenAI’s agent executed over 17,000 “attacker actions,” as referenced in Hugging Face’s interim technical report. The agent accessed four logins online that allowed entry to four separate, unnamed services. The letter identifies the models as GPT-5.6 Sol and an unreleased variant deemed even more capable.

      The term Alabama is absent from any of the requests,