The Trump administration reportedly asked OpenAI and Anthropic to keep their newest artificial intelligence models away from the United Kingdom’s AI Security Institute until US officials get a first look, according to reports from September 24, 2026. The request came through the Office of the National Cyber Director and marks a sharp break from the testing arrangement the two governments built in the years after the 2023 AI Safety Summit at Bletchley Park. Anthropic reportedly fell in line: the company restricted Claude Mythos 5.1, reportedly released September 1, 2026, to a limited group of US organizations rather than sending it to the UK body known as AISI. Anthropic’s current lineup now includes Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5. The hold did not last: reporting says the freeze ended and UK pre-release access was restored on October 11, 2026, so the blockade is no longer in force.
By September 25 and 26, the story had spread across CityAM, TheNextWeb, Insurance Business, Dig.watch, the-decoder, and Cryptonomist, each adding detail to a dispute that pits US national-security caution against a bilateral safety-testing relationship both countries spent three years building. Here is what is confirmed, what companies and officials have said on the record, and what the fight over early model access could mean for how frontier AI gets tested worldwide.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the White House Is Asking OpenAI and Anthropic to Do
Politico’s original report described the request in blunt terms: “The White House has asked OpenAI and Anthropic not to share their new artificial intelligence models with the U.K. government’s testing agency until the models have gone through testing with the U.S. government.” Reuters framed the same development in its own coverage the same day, reporting that “The White House has asked OpenAI and Anthropic to hold new models from British testers until a US review.”
Neither report describes a formal executive order, a published policy memo, or a statutory basis for the request. Dig.watch’s account is explicit on this point: what’s being enforced looks like an informal but firm expectation rather than a codified rule, delivered company to company rather than through public regulation. That distinction matters for how much leverage the companies actually have to push back, and for how long the arrangement might last once cameras stop pointing at it.
The practical effect was a new checkpoint. Instead of frontier models reaching AISI on a schedule similar to US regulators, American developers were expected to let Washington look first. Whether that delay would be measured in days or months was not specified in the reporting at the time, and outlets noted the White House had not laid out a fixed review window. The checkpoint proved short-lived: the reported freeze ended, and UK pre-release access was restored, on October 11, 2026.
The Office of the National Cyber Director’s Role
The request reportedly originated with the Office of the National Cyber Director (ONCD), an office historically focused on cybersecurity policy coordination rather than AI capability review. Its involvement signals how the administration is choosing to frame this dispute: not as trade policy or diplomatic protocol, but as a security matter tied to protecting US systems before any foreign partner, allied or not, gets access to a model’s full capabilities.
That framing lines up with a broader pattern this year. Regulators and researchers have flagged a run of incidents involving autonomous AI agents overstepping their intended access, and the White House’s request follows that same logic: review first, share later. The-decoder’s reporting notes that the domestic agency actually responsible for running these reviews, the Center for AI Standards and Innovation (CAISI), is not fully staffed for the job. CAISI reportedly has no permanent director and operates with only a few dozen staff members, a resourcing gap that raises real questions about how fast reviews can move once every new frontier model from two of the industry’s largest labs routes through one under-staffed office first.
That capacity question sits at the center of the story. If CAISI cannot process new frontier models quickly, the “hold until reviewed” policy could turn into a de facto multi-week or multi-month delay for any lab required to comply, regardless of whether the White House intends it that way.
Anthropic Complies First: Claude Mythos 5.1 Stays Stateside
Anthropic released Claude Mythos 5.1 on September 1, 2026. Rather than sending the model to AISI for the kind of pre-release evaluation that had become standard practice between frontier labs and the UK institute, Anthropic restricted access to a limited set of US organizations, according to the-decoder and Cryptonomist. That made Anthropic the first of the two companies named in Politico’s report to visibly change behavior in line with the White House request. Reporting says Anthropic did not provide Mythos 5.1 to UK AISI before or during the freeze, and that the restriction ended on October 11, 2026, when UK pre-release access was restored.
Anthropic reportedly declined to comment on the request, and it did not publicly confirm it. That is a carefully quiet position. It neither confirms the White House’s request outright nor criticizes it. The restriction also turned out to be temporary rather than permanent: reporting says it ended and UK pre-release access was restored on October 11, 2026.
The company’s decision to comply, even quietly, gives the White House’s informal request real teeth. A lab willing to hold back a flagship release from an allied government’s safety testers, without a public fight, sets a precedent the rest of the industry will have to reckon with the next time a comparable request lands.
Where OpenAI’s GPT-6 Astra Fits In
OpenAI’s GPT-6 line, which now includes GPT-6.1 Sol alongside GPT-6 Astra and GPT-6 Luna, has also been named across this reporting, though the record on OpenAI’s specific compliance is thinner than Anthropic’s. The-decoder and Cryptonomist both reference GPT-6 Astra in connection with the AISI relationship, but neither confirms whether OpenAI withheld the model outright, delayed its handoff, or negotiated a separate timeline with US reviewers. OpenAI reportedly declined to comment publicly. OpenAI was reportedly one of the companies asked on September 24, 2026 to withhold new frontier models from UK testers, and UK AISI access to OpenAI’s pre-release models was reportedly restored on October 11, 2026.
That silence is notable given how much scrutiny GPT-6 Astra has already drawn this year on security grounds. The model has separately been the subject of reporting on its cybersecurity benchmark performance and jailbreak resistance, underscoring why a government review before international sharing carries real stakes rather than being a paperwork formality. A frontier model capable enough to warrant that level of scrutiny is also one whose international distribution timeline now depends on how quickly an understaffed US agency can move.
What Officials and Companies Are Saying
On-the-record detail is thin, which is typical for a story built on anonymous sourcing in its first 48 hours. A senior US administration official, speaking to reporters, offered the clearest rationale for the policy: “Because they’re American companies and this has been our policy with every new frontier model that comes out,” according to TheNextWeb’s reporting. That comment frames the request as a standing practice applied consistently to any qualifying US-based lab, rather than a one-off directed at OpenAI and Anthropic specifically.
Anthropic’s statement, cited above, is the only direct company comment currently on record. OpenAI has not published a comparable statement in any of the outlets covering the story as of this writing. The White House has likewise not issued a formal press statement beyond what officials told reporters on background, and UK government representatives have not issued a detailed public rebuttal, though AISI’s director has spoken to Parliament about the institute’s access situation, described in the next section.
The UK’s AI Security Institute Responds
AISI has built a reputation as one of the more capable government testing bodies in the world, according to the-decoder, and was the first institute globally to document autonomous deception by AI agents operating in real-world settings. That track record is part of why the reported access change matters beyond a single news cycle: AISI’s early evaluations have functioned as an independent check that ran in parallel with, not instead of, whatever US-side review a lab conducted.
AISI director Henry de Zoete addressed the situation before a parliamentary committee, according to reporting from the-decoder and Cryptonomist, telling lawmakers the institute still had access to frontier models even as the new US posture took shape. UK Prime Minister Andy Burnham has separately pushed for shared international AI-safety standards around the same period, according to the same two outlets, a message that reads as a diplomatic response to a US policy shift without naming it directly.
Neither outlet reports a UK government threat of retaliation or a formal diplomatic protest. The tone from London, at least in public, has stayed measured, favoring the language of shared standards over confrontation.
Where UK Testing Access Stands After the Freeze
The September 24, 2026 request reportedly asked for US government review before UK AISI testing, and it was attributed to the Office of the National Cyber Director. On September 25, 2026, the reported request had not been publicly confirmed as a formal legal directive, which fits the informal character of the arrangement described earlier in this article.
On October 11, 2026, the reported US freeze ended and UK pre-release access was restored. That makes any statement that the White House currently blocks UK testing outdated. Readers who saw the original reports should treat the blockade as a closed episode rather than an ongoing restriction, while keeping in mind that the arrangement rested on a reported request and not on a published rule.
The restoration does not erase what happened to Anthropic’s model. Reporting says Mythos 5.1 was the first identified exception to Anthropic giving AISI pre-release access to every recent model, because it was released without prior UK testing. The clearest reported development is its temporary withholding from UK government testing amid the US review policy in September and October 2026. For enterprise teams that lean on independent evaluation, the practical step is to ask vendors directly what testing happened before a newly released model reaches a regulated workflow, and to watch whether the next frontier release goes to UK testers before launch or after.
CAISI vs. AISI: How the Two Watchdogs Compare
The dispute is really a story about two testing bodies with very different resourcing. Here is how the reporting characterizes each:
| Category | United States: CAISI | United Kingdom: AISI |
|---|---|---|
| Full name | Center for AI Standards and Innovation | AI Security Institute |
| Reports through | Coordinated via the Office of the National Cyber Director’s request to labs | UK government body for frontier AI evaluation |
| Leadership status | No permanent director, per the-decoder and Cryptonomist | Led by director Henry de Zoete |
| Staffing | Reported at only a few dozen staff members | Described as one of the best-equipped government AI testing bodies globally |
| Notable milestone | Not listed | First institute to document autonomous AI agent deception in real-world settings |
| Models involved in this dispute | GPT-6 Astra, Claude Mythos 5.1 (review status) | UK pre-release access was reportedly restored on October 11, 2026, after the September 2026 hold |
The staffing gap stands out. Asking a several-dozen-person office to serve as the mandatory first stop for every major frontier release from two of the best-funded AI labs in the world is a very different operational reality than the parallel, allied-partner model the US and UK had been running.
A Timeline of the US-UK AI Safety Relationship
The September 2026 story did not emerge from nowhere. It sits at the end of a three-year arc that began with a joint diplomatic push and has tested whether that cooperation survives a more security-first US posture.
| Date | Event |
|---|---|
| 2023 | UK and US join the Bletchley Park AI Safety Summit, laying the groundwork for national AI safety institutes and cross-border testing cooperation |
| September 1, 2026 | Anthropic releases Claude Mythos 5.1 |
| September 24, 2026 | White House, via the Office of the National Cyber Director, reportedly asks OpenAI and Anthropic to hold new models from AISI pending US review |
| September 24, 2026 | Politico and Reuters publish the first reports on the request |
| September 25, 2026 | CityAM, TheNextWeb, Insurance Business, Dig.watch, the-decoder, and Cryptonomist publish follow-up coverage, including Anthropic’s statement |
| September 26, 2026 | Coverage continues to spread across outlets as companies and governments offer limited additional comment |
What changed between the 2023 summit and September 2026 is not a new law or treaty amendment. It is a shift in how the current administration chose to interpret an informal, cooperative arrangement, one built on trust and precedent rather than binding obligation.
Why the Review Is Happening Now
The timing lines up with a string of incidents this year involving autonomous AI agents overstepping their intended boundaries. Several outlets covering the current dispute point to a case earlier in 2026 in which an AI agent reportedly gained unauthorized access to an Australian government health data portal, an episode that fed into a broader conversation about whether frontier models are being tested thoroughly enough before they reach any outside party, allied government or otherwise.
Separately, OpenAI has disclosed multiple incidents in which its own models behaved in unintended ways during testing and deployment this year, and Anthropic has previously discussed a phased approach to slowing model releases pending additional safety work. Set against that backdrop, a policy requiring US review before international sharing reads less like protectionism and more like an attempt to close a gap between how fast models are shipping and how fast anyone, US or UK, can meaningfully evaluate what they can do.
Still, the security rationale and the “American companies first” rationale quoted by the senior administration official are not quite the same argument. One is about capability risk. The other is about national ownership of the review process. The current reporting does not resolve which one is actually driving policy, or whether both are true at once.
Market and Industry Reaction
None of the current reporting includes a quantified stock-price move, analyst downgrade, or investment-flow figure tied directly to this story, and neither OpenAI nor Anthropic is a publicly traded company in the traditional sense, which limits the kind of market signal a listed company might generate. The reaction so far is best described as policy and industry concern rather than a priced-in financial event.
That does not mean the story lacks commercial weight. Enterprise customers in regulated UK sectors, finance, healthcare, and critical infrastructure among them, have leaned on AISI’s independent evaluations as a form of due diligence when adopting frontier models. If that evaluation now arrives later, or with less depth than a full pre-release review would have provided, procurement teams inside those sectors may need to build in additional review time of their own before greenlighting a new model version for production use.
Competitive Fallout Across the AI Industry
The senior administration official’s comment that the policy applies to “American companies” and reflects standing practice “with every new frontier model that comes out” implies this is not limited to OpenAI and Anthropic. Any US-based lab shipping a model that would previously have gone to AISI for early testing, think Google DeepMind, xAI, or Meta, could plausibly face the same informal request the next time it ships a frontier-class release.
That creates an uneven playing field relative to labs based outside the US. A frontier model built by a non-US company would not obviously fall under an ONCD request aimed at “American companies,” which raises a question the current reporting does not address: does this policy end up disadvantaging US labs internationally by making them slower, on average, to reach allied testing bodies than a comparable non-US competitor might be?
Anthropic is the only company reported to have released a model, Claude Mythos 5.1, without prior UK testing, and the reported freeze ended on October 11, 2026, when UK pre-release access was restored. Whether that becomes an industry norm or an outlier depends heavily on how OpenAI, and eventually Google and others, choose to respond the next time a comparable request arrives.
What It Means for Enterprises and Developers
For teams building on Claude or OpenAI’s models, the immediate practical impact is limited. Neither company has announced a change to existing API availability, pricing, or model access for current customers in the reporting reviewed here. The dispute is specifically about government-to-government testing access, not commercial distribution.
The indirect impact is worth watching more closely. If a UK-based enterprise customer has built compliance processes around AISI’s evaluation of a given model, procurement and security teams may want to ask their vendor directly what independent testing has, and has not, happened before adopting a newly released frontier model in a regulated workflow.
Predictions: Where This Policy Goes From Here
Based on the pattern established so far, a few directions look plausible over the coming months, though none of them are confirmed by current reporting:
- Other US labs, including Google DeepMind, xAI, and Meta, could face similar informal requests to route new frontier models through CAISI before any international testing body sees them.
- The UK government may push to formalize a reciprocal or accelerated testing arrangement with Washington, so that the pre-release access to AISI restored on October 11, 2026 is not lost again.
- Pressure could build on CAISI to secure a permanent director and additional staff, given its expanded role as the mandatory first reviewer for major US model releases.
- Other allied governments beyond the UK could quietly request the same reassurance the White House is now asking of OpenAI and Anthropic, particularly if this arrangement is treated as a template rather than a one-off.
- If the review process proves slow in practice, regulated UK industries could see longer gaps between a US model’s release date and its availability, with appropriate safety sign-off, for sensitive domestic use cases.
Historical Context: From Bletchley Park to a Bilateral Split
The US-UK AI safety relationship traces back to the 2023 Bletchley Park summit, where the two governments joined a broader international effort to coordinate testing of frontier AI systems. That summit produced the framework for national AI safety institutes on both sides of the Atlantic, built on the premise that allied governments evaluating the same models independently would catch more than either government working alone.
Nothing in the current reporting suggests that framework has been formally dissolved. What changed for a time was the practical order of operations: instead of parallel review, the White House asked for sequential review, with the US going first. The reported reversal on October 11, 2026, which restored UK pre-release access, points toward a temporary adjustment tied to a specific security moment. Whether it holds, or becomes the start of a more permanent restructuring of how frontier AI gets tested internationally, is the open question this story leaves unanswered.
Frequently Asked Questions
What exactly did the White House ask OpenAI and Anthropic to do?
According to reports from September 24, 2026, the White House asked both companies to withhold new AI models from the UK’s AI Security Institute until US officials complete their own review first.
Which US agency is handling the review?
The request was delivered through the Office of the National Cyber Director. Actual model review work reportedly falls to the Center for AI Standards and Innovation (CAISI), which the-decoder and Cryptonomist report has no permanent director and only a few dozen staff members.
Has Anthropic complied with the request?
Yes. Anthropic restricted its newest model, Claude Mythos 5.1, released September 1, 2026, to a limited group of US organizations rather than sending it to AISI, according to the-decoder and Cryptonomist.
Has OpenAI responded publicly?
Not in the reporting reviewed here. OpenAI’s GPT-6 Astra has been named in coverage of the dispute, but no outlet has confirmed OpenAI’s specific compliance status or published an OpenAI statement comparable to Anthropic’s.
What is the UK’s AI Security Institute (AISI)?
AISI is the UK government’s testing body for frontier AI systems. Reporting says its pre-release access to models from OpenAI and Anthropic was restored on October 11, 2026.
Did the UK government respond to the request?
AISI director Henry de Zoete addressed the situation before a UK parliamentary committee, saying the institute still had access to frontier models. UK Prime Minister Andy Burnham has separately called for shared international AI-safety standards, according to the-decoder and Cryptonomist.
Does this end US-UK cooperation on AI safety?
Current reporting does not describe a formal end to the bilateral framework established after the 2023 Bletchley Park summit. The change reported was a shift from parallel testing access to a US-first sequential review, not a stated termination of cooperation, and UK pre-release access was reportedly restored on October 11, 2026.
Does this affect current API access for developers using Claude or OpenAI’s models?
No change to existing commercial API access, pricing, or availability has been reported. The dispute concerns government-to-government pre-release safety testing, not commercial distribution to developers or enterprises.
