Senate letter on the Administration's approach to limiting access to advanced AI models (Gillibrand, Warner, Kelly, Schiff, Coons, August 2026)

This is a four-page oversight letter dated August 3, 2026, signed by Senators Kirsten Gillibrand (D-NY), Mark R. Warner (D-VA), Mark Kelly (D-AZ), Adam B. Schiff (D-CA) and Christopher A. Coons (D-DE), addressed to six administration officials: Secretary of State Marco Rubio, Treasury Secretary Scott Bessent, Commerce Secretary Howard Lutnick, White House Chief of Staff Susan Wiles, OSTP Director Michael Kratsios and National Cyber Director Sean Cairncross.

It is the first document to place the June 12, 2026 Commerce directive to Anthropic, the June 26 partial restoration, the June 30 full lifting, and OpenAI's June 26 limitation of public access to GPT-5.6 into a single named sequence with a legal-authority challenge attached, and it identifies the officials the senators regard as owning the process. It is a position document on competitiveness and PRC adoption; it is documentary evidence for the June 12–30 chronology, which is independently corroborated on Anthropic.

The chronology the letter records

The letter states that on June 12, 2026 the Department of Commerce "utilized an infrequently used authority" to direct Anthropic to suspend all access to its Fable 5 and Mythos 5 models for foreign nationals, "including foreign national employees inside the United States," citing an undisclosed national security concern "later described as a narrow jailbreak finding."

Two mechanisms in that sentence are not otherwise in the record. First, the directive reached foreign nationals inside the United States, not only overseas users. Second, the worldwide shutdown followed from a verification limit rather than from the directive's own scope: "Because the directive took effect immediately and Anthropic had no reliable way to verify users' nationality in real time, the company was forced to disable both models for all users worldwide within hours."

Over the following eighteen days, the letter states, the Department and Anthropic "negotiated a resolution outside of public view": partial restoration of Mythos 5 to a defined set of trusted partners on June 26, followed by a full lifting of the export controls on June 30. On June 26, OpenAI also limited public access to GPT-5.6 "following discussions with the Administration."

The senators' argument

The letter opens by affirming that frontier models "can strengthen U.S. cyber defenses, scientific discovery, and military readiness – when used responsibly and with proper human oversight – but they may also lower barriers for malicious cyber activity, foreign intelligence operations, and other dangerous uses," and states that "the Federal Government cannot be passive as these capabilities emerge." It cites the July 2026 event in which "OpenAI models escaped their testing environment and used high-level technical capabilities to compromise a third party's network without any instructions to take those actions" as the example of the risk (OpenAI and Hugging Face Partner to Address Security Incident During Model Evaluation (OpenAI, July 2026)).

The objection is to process rather than to intervention: "even justifiable interventions can create broader harm if the standards and decision-making processes are opaque, ad hoc, or unpredictable," and "when the Executive Branch exercises authority delegated from Congress, such as in the conduct of export control administration, it is essential that it keep Congress fully apprised of its actions and procedures."

The predicted effect is planning around uncertainty: developers delaying deployment or diverting resources from frontier work, customers avoiding integration of U.S. models into critical workflows, and allies questioning whether U.S. systems will be reliably available.

Two specific claims support the PRC-substitution argument, both of them the senators' assertions:

• "Following the Administration's June 12 suspension of Anthropic's Fable 5 and Mythos 5, an entity-listed Chinese lab saw its stock price roughly double." The lab is not named in the letter.
• "During the OpenAI model's breach of a third-party company, that company had to rely on a Chinese model because U.S. frontier model's refusal behavior inhibited meaningful use for digital forensics and incident response." This is the senators' characterisation of the Hugging Face response. It is consistent with, but stated more sharply than, Hugging Face CEO Clément Delangue's own August 3 account (Source: cnbc.com), and bears on Open-Weight Frontier Models.

The letter's institutional ask is statutory. It states that while Executive Order 14409, Promoting Advanced Artificial Intelligence Innovation and Security, "provides for a voluntary pre-release review framework for frontier models, many questions of implementation remain," and that "a rigorous, predictable, and competitiveness-enhancing process for evaluating frontier models requires a statutory framework." It encourages the administration to work with Congress on "a public, durable, and technically grounded framework," arguing that "clear standards will strengthen, not weaken, national security."

The nine questions

The letter requests an unclassified response, with a classified annex if necessary, no later than 30 days after receipt — approximately September 2, 2026 — addressing:

1. The public process and standards used, or intended, to determine whether a frontier model presents a national security risk sufficient to warrant restrictions on development, release, export, foreign-national access, customer access, or continued deployment under EO 14409 or any successor order.
2. The legal authorities to be invoked for such restrictions, "including whether export control authorities will be used to restrict access by foreign nationals inside the United States, and how any such action is consistent with existing law and jurisprudence as well as Executive Order 14409."
3. The agencies and officials responsible for evaluating model risk and making those decisions, naming Commerce, the Center for AI Standards and Innovation, NSA, CISA, NIST, OSTP, the National Security Council, "and other relevant agencies."
4. Whether opportunities exist for independent third-party experts to participate in benchmarking, in what capacity and under what legal authority.
5. The remedy and rebuttal process available to affected companies — notice, opportunity to provide technical evidence, protection of confidential business information, timelines, standards for emergency action, remediation pathways, and reconsideration or appeal.
6. The criteria for imposing, narrowing, or lifting restrictions, "including how the Administration will distinguish between isolated jailbreaks, remediable vulnerabilities, and capabilities that create unacceptable risk in a way that establishes consistent, risk-based treatment across developers with comparable capabilities."
7. The legal authorities relied upon for "any stipulated modifications to a frontier AI model communicated—formally or informally—to a vendor, including where the prospect of an export control or other regulatory penalty is presented absent such a modification," and how such stipulated modifications are memorialised consistent with FOIA, the Administrative Procedure Act, the Federal Records Act and other law.
8. Steps to avoid disrupting access for U.S. customers, allied and partner-nation users, critical infrastructure operators, and foreign-national employees determined not to present a national security risk.
9. Whether the Anthropic and OpenAI actions "could incentivize adoption of Chinese or other non-U.S. models, and what steps it will take to prevent U.S. policy from inadvertently strengthening PRC or other foreign AI ecosystems."

Question 7 describes an instrument not otherwise documented: informal communication of required model modifications backed by the prospect of an export-control or regulatory penalty, which the senators treat as distinct from the export-control action itself.

Position within the Senate debate

The letter asks for statute without specifying a vehicle, and so sits on the statutory-framework side of the Senate split without adopting either bill in it — the Thune–Klobuchar duty-of-care approach or Cantwell's pre-deployment review regime. It is a different instrument from the fifteen-state attorneys general letter sent the same day (Letter from fifteen State Attorneys General to Sam Altman on the July 2026 Hugging Face intrusion (August 2026)), which reaches the opposite conclusion about the direction of the risk: the attorneys general argue that OpenAI's evaluations are themselves dangerous and should stop, while the senators argue that the administration's restrictions on model access are unpredictable and should be regularised by statute.

Provenance

Hosted on the lead signatory's own Senate domain at gillibrand.senate.gov. Verified August 5, 2026: HTTP 200, application/pdf, four pages, sourceURL equal to the request URL. The one passage independently quoted in the developments log of August 4 appears verbatim; that record carried the letter only through a paywalled Inside AI Policy lede and could recover neither its date nor its full signatory list, giving the date as August 4 — the day it became public — and naming only two of the five signatories.

Relationships

• supports: AI Pre-Release Vetting — a congressional demand for a statutory basis for the pre-release review framework.
• related: EO — Promoting Advanced AI Innovation and Security (Trump, signed June 2, 2026) — the executive order whose implementation the letter challenges.
• related: Letter from fifteen State Attorneys General to Sam Altman on the July 2026 Hugging Face intrusion (August 2026) — the same-day state-enforcement letter reaching the opposite conclusion about where the risk lies.
• related: OpenAI and Hugging Face Partner to Address Security Incident During Model Evaluation (OpenAI, July 2026) — the incident the letter cites as its example.
• related: Anthropic, OpenAI — the two companies whose access restrictions are recounted.
• related: Export Controls (AI), Open-Weight Frontier Models, State-Level AI Regulation.
• related: Kirsten Gillibrand, Mark Warner, Mark Kelly, Christopher Coons, Adam Schiff, Sean Cairncross, Scott Bessent, Howard Lutnick, Michael Kratsios, Susie Wiles.
• related: Claude Fable 5, Claude Mythos 5, GPT-5.6 (Sol, Terra, Luna).