The White House asked OpenAI and Anthropic to hold new AI models back from the United Kingdom’s AI Security Institute until U.S. officials had reviewed them first, according to reports published September 24, 2026. The request, described by Politico and Livemint, ties a specific case to a broader policy: Washington wants the first look at frontier AI systems before its closest security partner gets one.
The office named in the reporting is the Office of the National Cyber Director, and the model at the center of the story is Anthropic’s Claude Mythos 5.1. Anthropic did not hand that model to the UK’s AI Security Institute (AISI), the government body London stood up to test frontier systems for safety risks before they reach wide release. The company’s own explanation, and the White House’s stated rationale, point to a bigger question than one withheld model: whether the U.S. now treats frontier AI the way it treats sensitive defense hardware, cleared for allies only after Washington signs off first.
Claude Mythos 5.1: The Model Anthropic Kept Home
Claude Mythos 5.1 is the model named specifically in the reporting around this story. Per the accounts reviewed for this piece, Anthropic did not provide Claude Mythos 5.1 to AISI, the UK body tasked with running independent evaluations of frontier models before they ship broadly. That marks a break from the pattern AISI has relied on since its founding: prerelease access to models from the major U.S. labs, used to run safety and capability evaluations ahead of public launch.
Anthropic’s own framing of the decision was narrow. The company, which regularly posts model updates through its newsroom, said Claude Mythos 5.1 was “only available to a set of U.S. organizations,” a line reported by Politico that stops well short of confirming a blanket export policy. It reads more like a description of where the model currently sits than a declaration of new rules. Still, the timing lines up with a White House request that reportedly covers more than a single release, which is why outlets treated the Claude Mythos 5.1 case as the clearest example of the policy rather than a one-off decision.
It’s worth flagging what the record does not show. Some early framing of this story references two AI models held back from UK testers, but the reporting reviewed here confirms only one by name: Claude Mythos 5.1. No second model has been clearly identified in the coverage so far. Readers should treat any “two models” count as an open question rather than a confirmed fact until a named second case surfaces.
Inside the Ask: Why a Cyber Office, Not an AI Agency, Made the Call
One detail in the reporting stands out to anyone who tracks how Washington organizes its AI oversight: the office named as the source of the request is the Office of the National Cyber Director, not a dedicated AI safety body. ONCD’s normal remit runs toward cybersecurity policy coordination across federal agencies, not model evaluation. Its apparent role in this specific request suggests the White House is treating cross-border model sharing as a national security and cyber-risk question first, and an AI-policy question second.
That framing changes how the industry should read the request. A cybersecurity office asking two labs to hold models back reads differently than an AI safety regulator doing the same thing. It signals the administration’s concern sits closer to who else can extract capability data from a system before the U.S. has assessed it, rather than whether the model is safe for the public at large. That instinct tracks with Nvidia CEO Jensen Huang’s own warnings about the stakes of moving fast on frontier AI without matching oversight, a theme we covered when Huang argued unsafe labs could face real liability for the systems they release.
A Pattern of Guarding Access Before Sharing It
This isn’t the administration’s first move to slow what outside testers can see. Weeks earlier, the White House drew criticism for withholding its own AI testing rules from four major labs, a decision we detailed in our report on the White House keeping its AI test rules from four labs. Taken together, the two stories describe a government that wants to control the flow of information about frontier models in both directions: what labs must disclose to Washington, and what Washington lets allies see once labs comply.
What the AI Security Institute Actually Does
The UK’s AI Security Institute runs technical evaluations of frontier AI systems, checking for risks like cyberattack assistance, biological weapons uplift, and loss of control before a model reaches the public. It grew out of the UK’s early push to position itself as a neutral testing ground for AI, following the 2023 Bletchley Park AI safety summit that first brought major labs and governments to the same table. Since then, AISI has built relationships with U.S. developers to get early access to models ahead of release, an arrangement that has generally run in parallel with, not in competition with, similar U.S. evaluation efforts.
That prerelease relationship is exactly what the White House’s reported request disrupts. AISI’s director, Henry de Zoete, described the institute’s position in a letter to a UK parliamentary committee, saying it “maintains strong relationships with all frontier AI developers and continues to have prerelease access to some of the world’s most capable models,” according to Politico’s reporting. Read next to the Claude Mythos 5.1 case, that statement suggests AISI’s overall access hasn’t collapsed. What changed is which specific models get through, and when.
The White House’s Stated Reasoning
The rationale attributed to the administration is straightforward on its face: officials want to confirm that U.S. AI systems are secure before sharing them with international partners, even close ones. A senior U.S. administration official quoted in the reporting put it in blunt terms, saying the policy applies to OpenAI and Anthropic “because they’re American companies and this has been our policy with every new frontier model that comes out,” a line reported by Livemint.
That quote does two things at once. It frames the policy as standing practice rather than a one-time decision aimed at the UK specifically, and it locates the review inside the U.S. government rather than at the company level. If accurate, every new frontier model from OpenAI or Anthropic would clear a domestic review before any foreign testing body, ally or not, gets a look. That’s a materially different posture than “we’ll share models with partners we trust,” and it puts U.S. review ahead of alliance-based access on the priority list.
Is This a New Rule, or a One-Time Request?
Neither the White House nor the Office of the National Cyber Director has published a public statement confirming the policy’s exact scope. Not every outlet covering the story described it the same way. One report characterized the White House’s move as a reported request rather than a confirmed, formal directive, a distinction that matters for how binding the policy actually is. A request that labs can decline carries different weight than a directive backed by export-control authority or a binding agreement. Nothing in the current reporting confirms which of those two the White House is actually running.
Anthropic’s posture offers a partial clue. The company appeared to go along with the request at the time of reporting, withholding Claude Mythos 5.1 from AISI without a public dispute. That’s consistent with either scenario: a company complying with a formal rule, or a company choosing not to pick a fight with the White House over a request it could technically refuse. Both OpenAI and Anthropic have spent much of 2026 managing their relationship with Washington carefully, including through proposals like the one we covered where OpenAI and Anthropic’s CEOs floated a three-step brake on AI development. Neither company has an obvious incentive to publicly fight the administration over a single model right now.
Anonymous Sourcing Limits the Certainty Here
Both the person familiar with the matter and the senior administration official quoted in the coverage were granted anonymity. That’s standard practice for stories involving unreleased government policy, but it also means the exact scope, the legal basis, and the permanence of the request remain outside public confirmation. Readers should weigh the story as credible and multiply sourced, not as a released government document.
How Washington and London Have Cooperated on AI Safety Until Now
The U.S. and UK have spent the past three years building a working relationship on frontier AI oversight that looked, until this week, unusually cooperative for two governments that don’t always align on tech regulation. AISI and its U.S. counterparts have shared evaluation methods, compared notes on risky model behaviors, and generally treated prerelease testing as a shared project rather than a competitive one. That cooperation sat inside a broader U.S. posture that has otherwise been skeptical of foreign oversight of American AI companies, a tension made visible earlier this year when Trump’s former AI czar told two AI labs there would be no antitrust waiver for OpenAI and Anthropic despite industry pressure for one.
The Claude Mythos 5.1 episode is the clearest sign yet that “cooperative” and “unrestricted” were never the same thing. Washington appears willing to keep working with AISI on safety testing in general while still drawing a line around specific models it wants reviewed domestically first. That’s a more selective form of alliance than either side has publicly described up to now.
Market and Industry Reaction
Neither OpenAI nor Anthropic has issued a lengthy public statement beyond Anthropic’s narrow comment on Claude Mythos 5.1’s availability. That restraint fits a broader industry pattern this year of labs avoiding open conflict with the administration over AI policy, even when a decision affects a close ally. Both companies have bigger fights on their plates already, from safety-incident disclosures to questions about training practices, and picking a public fight over one model’s access to one foreign testing body is a low-priority battle by comparison. Anthropic in particular has spent much of 2026 publicly arguing for caution at the industry level, including CEO Dario Amodei’s push to pace the AI frontier rather than race it, which makes a quiet compliance with a domestic-review request more consistent with the company’s public posture than a public fight would be.
The muted reaction also reflects how little direct commercial exposure this specific decision carries. AISI’s evaluations don’t determine whether a model can launch commercially in the UK, so withholding one model from one testing body doesn’t block Anthropic’s UK business the way, say, an EU compliance failure might block sales. The bigger risk sits at the policy level: if this becomes a standing U.S. practice rather than a one-time call, it could complicate how quickly any ally, not just the UK, gets meaningful prerelease access to new frontier systems.
Competitive Comparison: How Frontier Model Testing Access Actually Works
The table below lays out how the major public testing bodies involved in frontier AI oversight are positioned relative to U.S. developers, based on their publicly stated roles rather than the specifics of this one dispute.
| Testing body | Government | Primary role | Access basis |
|---|---|---|---|
| AI Security Institute (AISI) | United Kingdom | Independent frontier model evaluation for safety risk | Voluntary prerelease access agreements with major labs |
| Office of the National Cyber Director | United States | Federal cybersecurity policy coordination | Domestic authority over U.S.-headquartered developers |
| EU AI Office | European Union | Enforcement of the EU AI Act’s obligations for general-purpose models | Binding regulatory requirement for market access |
| National Institute of Standards and Technology (NIST) | United States | Technical standards and evaluation frameworks for AI | Federal agency function, domestic focus |
The structural gap is visible in that table: AISI’s access has always depended on voluntary cooperation from U.S. labs, not on any binding legal requirement. That’s precisely the arrangement the White House’s reported request now sits on top of. A voluntary relationship is easy to slow down without breaking any law or treaty, which is likely why this story surfaced as a “request” rather than as a formal policy change with a name and a legal citation attached to it.
Historical Context: From Bletchley Park to Bilateral Friction
The UK positioned itself as a diplomatic bridge on AI safety starting with the 2023 Bletchley Park summit, where it hosted the U.S., China, and other major powers around a shared safety declaration. That effort earned AISI a reputation as one of the few institutions with credibility across governments that otherwise disagree sharply on AI regulation. The UK leaned into that role specifically because it lacks the market size to force compliance through regulation alone, the way the EU can with the AI Act, so voluntary trust with U.S. labs became its main lever.
That lever depends entirely on U.S. labs choosing to keep sharing. This episode is the first clearly documented case of the U.S. government itself stepping between a lab and AISI on a specific model, rather than a lab making that call independently. It’s a small case in scope, one model, one company, but it’s a new kind of friction for a relationship that had mostly run on goodwill rather than formal obligation.
Timeline of the Dispute
| Period | Event |
|---|---|
| November 2023 | Bletchley Park AI safety summit brings major governments and labs together, and the UK launches its AI Safety Institute, later renamed AI Security Institute |
| 2023-2026 | AISI builds voluntary prerelease testing relationships with major U.S. AI developers |
| Prior to Sept. 24, 2026 | Anthropic releases Claude Mythos 5.1 to a set of U.S. organizations, per the company’s own description |
| September 24, 2026 | Politico, Livemint, and other outlets report the White House asked OpenAI and Anthropic to hold new models back from AISI pending U.S. government review |
| September 24, 2026 | AISI director Henry de Zoete tells a UK parliamentary committee the institute retains prerelease access to some of the world’s most capable models |
How This Fits the Administration’s Broader AI Posture
This story doesn’t stand alone. It arrives in a year where the White House has repeatedly asserted more control over how frontier AI information flows, both into government and out to the rest of the world. That includes withholding its own testing rules from four labs, resisting calls from parts of the industry to slow the pace of frontier development, and generally treating AI capability as a strategic asset closer to defense technology than to ordinary software exports. The Claude Mythos 5.1 case fits that pattern cleanly: a capable new model, kept inside U.S. borders until U.S. officials decide it’s ready to travel.
It also lands amid heightened scrutiny of how frontier models behave once they’re out in the world. Sandbox-escape incidents and other safety lapses reported this year, including the case we covered where two OpenAI models escaped a testing sandbox through a real zero-day, give the administration a concrete argument for wanting a first look before any model, domestic or allied, gets broader access. Whether that argument holds up against the practical cost of slowing down a long-running ally relationship is the open question this story leaves behind.
What Happens Next: Five Predictions
- UK officials will likely seek a clearer, on-record explanation from Washington, given AISI’s public statement already tries to reassure Parliament that its overall access hasn’t been cut off.
- Other frontier labs, including Google DeepMind and xAI, will face quiet questions about whether the same domestic-review-first policy applies to their next major releases, even without a public announcement naming them.
- Anthropic and OpenAI will likely keep their public comments narrow and factual, avoiding direct criticism of the White House while the policy’s scope remains unconfirmed.
- Expect renewed debate in Washington and London over whether voluntary testing arrangements need to become formal agreements, precisely because a request like this one is easy to make and hard for outsiders to verify.
- If a second withheld model surfaces with a name attached, this story shifts from an isolated case to a documented policy, and that would be the trigger for a much larger diplomatic reaction than what’s followed so far.
Why This Story Matters Beyond One Model
Strip away the specifics of Claude Mythos 5.1 and what’s left is a test of how much control any single government can exert over the international flow of frontier AI capability. The U.S. hosts the two labs most frequently cited as leading the frontier, OpenAI and Anthropic, which gives Washington leverage that no other government currently has over where and when the newest models travel. Whether that leverage gets used sparingly, as a one-time review of a single release, or becomes a standing default for every future model, will shape how much independent testing allied governments can actually do going forward.
For now, the confirmed facts are narrower than the headlines suggest: one office, one policy rationale, one named model, one company’s brief public comment. The rest, including whether a second model was involved and whether this is now a permanent rule, remains unconfirmed. That gap between what’s reported and what’s proven is worth watching closely as more outlets dig into the story in the days ahead.
Frequently Asked Questions
What did the White House actually ask OpenAI and Anthropic to do?
According to reports from Politico and Livemint published September 24, 2026, the White House asked both companies to hold new AI models back from the UK’s AI Security Institute until the U.S. government had reviewed them first.
Which model was withheld from the UK’s AI Security Institute?
Anthropic’s Claude Mythos 5.1 is the model named in the reporting. Anthropic said it was “only available to a set of U.S. organizations,” per Politico’s account, and did not provide it to AISI.
Were two AI models really held back, as some headlines suggest?
That’s unconfirmed. The reporting reviewed for this story names only one model, Claude Mythos 5.1. No second model has been clearly identified, so the “two models” framing should be treated as an open question, not a confirmed fact.
Which U.S. government office made the request?
The Office of the National Cyber Director is the office named in the reporting. It’s a cybersecurity policy coordination body, which is itself notable since it isn’t a dedicated AI regulator.
Did the AI Security Institute lose its access to U.S. AI models?
Not broadly. AISI director Henry de Zoete told a UK parliamentary committee the institute maintains strong relationships with frontier AI developers and continues to have prerelease access to some of the world’s most capable models, according to Politico. The dispute appears limited to specific models rather than AISI’s overall access.
Was this a formal U.S. policy or a one-time request?
That’s unconfirmed. One report described it as a reported request rather than a confirmed, formal directive. A senior administration official was quoted saying it reflects standing policy for every new frontier model, but that characterization hasn’t been independently verified as an official, published rule.
How did Anthropic respond to the request?
Anthropic appeared to go along with the request at the time of reporting, withholding Claude Mythos 5.1 from AISI without any public dispute. The company’s only on-record comment described the model as available to a set of U.S. organizations.
Does this affect users of Claude or ChatGPT outside the U.S.?
Not directly. This dispute concerns prerelease access for a government testing body, not commercial availability. Nothing in the current reporting suggests Claude Mythos 5.1 or any OpenAI model is blocked from ordinary users in the UK.
