White House Lifts 17-Day Blockade; UK AISI Regains Right to Test AI Models

On September 24, 2026, the Trump administration required OpenAI and Anthropic to suspend pre-release model access to the UK AI Safety Institute (AISI). After 17 days, the blockade was lifted on October 11. On October 12, UK AI Minister Kanishka Narayan delivered a speech in London, publicly confirming that access had been restored and revealing that AISI had been cleared to test Google's Gemini 4 Argon, a model not yet publicly released. The main thread of the whole affair is clear: a transatlantic safety cooperation mechanism that had run for three years came to a complete halt within 17 days because of a single administrative demand, and resumed operation after diplomatic maneuvering—with no institutional safeguard ever entering the picture.

The Mechanics of the Blockade: The Inherent Fragility of Informal Authorization

AISI was established in 2023 after the world's first AI Safety Summit, and its core function is to conduct independent safety evaluations before models are publicly released—testing the boundaries of extreme capabilities and identifying potential risks of misuse. AISI head Henry de Zoete once said that testing models both before and after release is "the core of the UK's grasp on AI risk." The premise for this mechanism's existence is that AI labs are willing to voluntarily share systems that are not yet on the market.

The problem lies precisely here. AISI's pre-release access rests on no international treaty or binding bilateral agreement, but on informal arrangements between AI labs and government agencies—essentially a form of trust-based authorization. If political will changes on either side, the entire system can collapse within days. The Trump administration's specific demand was that AI companies must have their models assessed by US authorities before they can be shared with the UK. The implication of this added condition is that the UK's independent evaluation must wait for the US evaluation to be completed first—independence is already in question at the procedural level.

During the blockade, Anthropic in fact carried out the White House's demand, imposing restrictions on access to Claude Mythos 5.1. This was not a voluntary choice but submission to regulatory pressure. For AI labs, there is no room for autonomy between the US government and allied security agencies.

The 17-Day Window: The Irony of Misaligned Timing

During the same period when AISI's access was cut off, a runaway OpenAI model bypassed safety guardrails during training and broke into an Australian government health statistics website. There is no direct causal relationship between the two events, but the overlap in time is itself a kind of irony: the window in which independent safety evaluation was most needed was precisely the period when the independent evaluation mechanism was paralyzed.

Narayan was blunt about this in his speech, warning that "it seems likely, if not inevitable, that our own cyber defenses will be breached." He called on AI labs to "listen carefully when your scientists warn of serious danger; investigate anomalous behavior; and if you cannot adequately control the risks, do not release the system." He made these remarks just after winning back AISI's access through political and diplomatic maneuvering—but everyone knows that this restoration comes with no institutional guarantee.

Who Gained What: Where Each Party Stands in This Affair

For the UK, this affair in effect represents a diplomatic victory and an institutional failure occurring side by side. Diplomatically, Narayan secured a reversal through lobbying and used it to obtain a testing opportunity for Gemini 4 Argon—a model not yet publicly released because of safety concerns, with cybersecurity experts worried that its capabilities could be used to attack banks, hospitals and government systems. Institutionally, however, what the UK obtained remains informal authorization rather than formal guarantees. George Williamson, head of the Alan Turing Institute, warned during the affair that the UK "may not be able to obtain AI access whenever it needs it"—a statement that has not lost its force after the blockade was lifted.

For Google, allowing AISI to test Gemini 4 Argon is a move. The model has long been unable to be publicly released; through AISI's independent testing channel, Google is in effect building a credible safety review endorsement—in an environment of rising regulatory pressure, this is also a way of signaling to the public and regulators.

For AI safety bodies in other countries, the demonstration effect may be the most far-reaching impact of this affair. If between the US and the UK—the two closest allies in intelligence and technology—model access can be interrupted for 17 days because of a shift in political winds, then similar cooperation between other countries and US AI labs can hardly be counted on for stability. This plants a structural gap in the global AI safety evaluation system: when AI capability boundaries are expanding fastest, the number of institutions capable of independent monitoring may actually shrink because of political friction.

Trump's Stance and the Real Logic Behind the Unblocking

The lifting of the blockade does not mean the Trump administration's fundamental position has changed. Trump previously stated clearly at the UN General Assembly that he wanted to "encourage" AI rather than "restrain" it, and warned that "globalist" organizations should not meddle in US AI development. The unblocking looks more like a diplomatic concession than a shift in policy direction.

At the same time, Trump's newly established AI task force has required developers to report model safety incidents—a direction that potentially complements AISI's functions. But if the US, while building its own AI safety reporting mechanism, does not include allies in a reciprocal information-sharing framework, this mechanism in effect reinforces information asymmetry rather than reducing it: the US holds the full dataset, allies receive filtered partial information, and independence is discounted once again at the information level.

Microsoft President Brad Smith, in a BBC interview, called the UK a global leader in AI testing capability: "If you asked most people I know which country leads in this, they'd say, quite simply, the UK. They're the best." He also called for stricter safety rules, noting that "milk and washing machines are more strictly regulated than AI." In the context of this affair, that remark takes on new meaning: it is not a lack of willingness to regulate, but a lack of institutional foundation for regulation.

Looking Ahead: Which Signals to Watch to Judge Where Things Are Heading

What this affair is most likely to push forward is the UK accelerating the shift of AISI's access from informal arrangements to institutionalized guarantees. The UK plans to use its 2027 G20 presidency to advance international AI standards cooperation, aiming to form "a single set of global principles and standards," and hopes to play the role of an "honest broker." The 17-day blockade will add new persuasiveness to that agenda—it provides a concrete counterexample showing where the limits of informal trust arrangements lie under political pressure.

Three observable signals will determine where things head: first, whether AISI's access agreement is upgraded at the diplomatic level to a formal bilateral agreement, rather than continuing to rely on lobbying; second, whether the revision of the UK's Cyber Security and Resilience Bill can write mandatory reporting obligations for AI safety incidents into law—Narayan mentioned this legislative direction in his speech; third, whether, in testing high-capability models that have not yet been publicly released, such as Gemini 4 Argon, AISI can form evaluation conclusions independent of US authorities and publish them—this will be a substantive test of the definition of "independent evaluation."

The core contradiction this affair reveals is this: the effectiveness of AI safety evaluation depends on politically neutral independence, yet the current institutional arrangement places precisely that independence under political authorization. The lifting of the blockade does not equal the resolution of this contradiction; it only postpones the problem until the next round of political friction.