Warren said the safety tests are hidden. She asked what happens when a model fails.

WASHINGTON — October 6, 2026

Tuesday’s new document is a letter, not a test. Senators Elizabeth Warren and Richard Blumenthal dated it October 5. Warren’s office published it Tuesday. Ashley Gold at Semafor, in a story time-stamped 4:54 a.m. Eastern, had the letter first. By evening the order they are asking about still does not say what happens if a model fails.

A request for the fail rule is not the fail rule. The page on the White House website still classifies the benchmarking, calls the access voluntary, and creates no right anyone can enforce.

The sentence on the page

The letter asks for the drafting history of Executive Order 14409 and for the testing process that order created. Warren’s release names six recipients: Treasury Secretary Scott Bessent, Commerce Secretary Howard Lutnick, White House chief of staff Susie Wiles, National Cyber Director Sean Cairncross, science adviser Michael Kratsios, and budget director Russell Vought. The PDF prints the chief of staff as Susan Wiles. The salutation names Bessent, Lutnick, Cairncross, Vought, and Ms. Wiles. Kratsios is on the address block. He is not in the greeting.

The deadline is printed. Answers “no later than October 19, 2026.”

The sentence Semafor shortened is the one to keep. The senators wrote that Americans deserve “strong guardrails set by their democratically elected officials, not a toothless framework shaped in secret by a handful of executives at self-interested trillion-dollar companies.” Gold’s story ends that line at “a handful of billionaires.” The PDF does not. A paraphrase is not the filing.

The sentence that governs the rest is the one about the tests. “At a moment that calls for independent, transparent, and enforceable safeguards, the American people have been left with an entirely voluntary, industry-designed oversight process, in which the implementation guidelines, safety tests, and decision criteria are hidden from the public.” Hidden is their word. It is an allegation that the criteria have not been shown. It is not a copy of the criteria.

Sixty days in the order is a deadline, not an exam

The document they are asking about is already public, and it is narrower than the draft they describe. Executive Order 14409, “Promoting Advanced Artificial Intelligence Innovation and Security,” is dated June 2, 2026. Section 3 gives the Treasury secretary, the secretary of war through the director of the National Security Agency, and the homeland security secretary through the director of the Cybersecurity and Infrastructure Security Agency sixty days from that date to do two things.

The first is “a classified benchmarking process to assess the advanced cyber capabilities of AI models and determine the threshold at which an AI model should be designated a ‘covered frontier model.’” The order says that determination shall be made by the NSA director, in consultation with the national cyber director and others. Classified is the order’s adjective. A letter from two senators does not declassify it.

The second is “a voluntary framework” through which developers would be able to ask whether a model meets the designation, give the government access “for a period of up to 30 days before they plan to release such models to other trusted partners,” and help choose those partners. Up to 30 days. Before trusted partners. Not a public score.

Section 3(c) refuses the other reading. “Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement” for the development, publication, release, or distribution of new AI models, “including frontier models.” Section 5(c) says the order “does not create any right or benefit, substantive or procedural, enforceable at law or in equity” against the United States. A voluntary window, and no private right, is what the page still says.

The letter’s account of how that page was reached is labeled as reporting, and it should stay labeled. Citing a September 20 Politico Magazine story by Sophia Cai and Cheyenne Haslett, the senators write that an Office of the National Cyber Director draft in May would have imposed a mandatory 60-day government evaluation before a limited release, plus an additional 75-day access period before a broader public release, with the government able to block a system that posed an unacceptable risk. They write that executives from Google, Anthropic, and OpenAI “responded negatively,” viewing the proposal as “too prescriptive in regulating model development, release timing and customer eligibility,” and that OpenAI complained the framework was “not a light-touch voluntary coordination mechanism.” They write that the mandatory draft then became voluntary and applied to a more limited set of models, and that executives from OpenAI and Microsoft praised Cairncross as a “capable leader” who is “committed to getting the balance right.”

Then a second cut, still on their citations. After Meta chief executive Mark Zuckerberg and xAI chief executive Elon Musk “reportedly expressed concerns to President Trump,” the administration “further watered down the framework from a 90-day testing window to 30-day window, which became Executive Order 14409.” The footnote is the Politico story, a May 22 piece in The Decoder, and the order.

Hold the numbers apart. Sixty plus seventy-five is not ninety, and the letter does not show the arithmetic. The 60 days in the signed order are a clock to build the process, counted from June 2. They are not an evaluation that lasts 60 days. The 30 days in the signed order are “up to 30 days” of access before release to trusted partners. The 60-day exam, the 75-day wait, and the 90-day window live in the letter’s account of what reporters were told about drafts. This desk has not been shown those drafts. The second numbered request in the letter is the request for them, “along with redlines specifying the substance and date of each material change.”

The question the order skips

There are five numbered requests, and a reply date of October 19. The first two ask for meetings and for communications with private-sector firms about the order and about drafts “as of early May, May 18, and May 21,” plus every draft and those redlines. The third asks for “a copy of the reported 31-item questionnaire given to industry firms by the ONCD,” the list of firms that got it, and their responses. Reported. The citation is again Politico. The form is not attached.

The fourth asks whether anyone outside the industry was consulted on the scope of the order or on “the classified benchmarking and evaluation process.” The examples are independent evaluators, civil society, consumer advocates, operators of critical infrastructure, and academic experts. If the answer is yes, they want the paper.

The fifth is the one that matters if a model is actually dangerous. A full description of the pre-deployment testing process: the criteria for designating a covered frontier model, the capabilities and risk domains being evaluated, and the benchmarks. The PDF this desk read does not letter the last clause as its own subpart. The clause is still the question. “What actions, if any, are taken if a model fails an evaluation or is deemed to pose an unacceptable level of risk.”

If any. The order’s voluntary framework describes access and consultation. It does not describe a fail. Section 4 tells the attorney general to prioritize criminal laws against anyone who uses AI to illegally access or damage a computer. That is a charge against a person using a tool. It is not a sentence that says a lab’s model, having failed a classified benchmark, does not ship.

October 9 is a different letter

This is not the first piece of paper. On September 28, Blumenthal and Warren wrote to Bessent alone. Blumenthal’s office published that letter the same day and set the reply at October 9. It said the benchmarking process was “reportedly finalized in August” and that there had been “no public information about the criteria, participation, or effectiveness of the program.” Its seventh question already asked whether any company “has failed or refused to address risks or vulnerabilities identified under the testing program.”

October 9 has not arrived. An open deadline is not a refusal. What Tuesday adds is a wider address list, the drafting record, the questionnaire, and a direct ask for the consequence of failure. The September letter’s questions about Hugging Face, escaped sandboxes, and government websites are not the new lead. The new lead is who wrote the voluntary page, and what the page does when a model does not pass.

The accord already on file here is a different document, and the new letter treats it as an example rather than as the order. The senators write that Zuckerberg played a “central role in shaping” the White House Accord on Super Intelligence, that he “circulated a draft,” and that Jensen Huang worked to gather other executives, citing Ashley Gold’s September 30 Semafor account. They quote the president’s “morally binding.” The lunch, and the four rules that named no penalty, are already public. So is the force the spy chief was given on Sunday. Tuesday’s letter does not replace either. It asks for the testing manual the June order said would be classified.

One claim should not be stretched across a probe this desk has already published. The senators write that the Federal Trade Commission under President Trump “has dismantled or vacated numerous existing investigations and enforcement orders” on unfair or deceptive uses of AI. The footnotes are the White House AI action plan and a February Reuters legal commentary. This letter does not mention the inquiry Chairman Andrew Ferguson’s Commission confirmed on September 30. A sentence about older orders is not a sentence about that file.

What Tuesday evening does not contain

No reply. Semafor’s story and the senators’ release do not print one, and this desk has not found an administration answer in the public record through Tuesday evening. Two senators writing from the minority did not attach a subpoena. October 19 is a date they set. It is not a contempt date.

No draft. The 60-day evaluation, the 75-day access period, the 90-day window, and the 31-item questionnaire are in the letter as reported. They are not exhibits. The signed order’s “up to 30 days” is an exhibit. It is voluntary, and it is access before trusted partners.

No new incident. The letter does not disclose a fresh breakout, a resignation, or a model that failed a government test and was stopped. It asks what the government would do if one did. The order, as posted, still does not say.