Anthropic’s AISI refusal heralds the end of the road for mutual collaboration on frontier AI assessment

Lead author: Prof Aled Lloyd Owen with contributions from Prof Gina Neff, Prof Gopal Ramchurn and Prof Dame Wendy Hall

Summary

Anthropic’s refusal to submit Mythos 5.1 to scrutiny by the UK AI Security Institute (AISI) deals a blow to the cooperative model upon which UK national security scrutiny of the most capable AI models hinges.  Anthropic’s decision reflects the weaknesses of the UK’s approach to AI regulation and the increasing tension between the voluntary collaboration of frontier AI companies with regulators, and the limitations of regulators’ power of compulsion.  This lack of cooperation, coupled with the alarming sentiments of Anthropic engineers about human extinction, could further damage public trust and confidence in AI when it is most needed to support beneficial adoption and uptake.

What AISI offers

AISI conducts cyber evaluations on new releases of the most capable autonomous systems. Its evaluation of Claude Mythos Preview in April 2026 received acclaim for its methodology and conclusions, including from Anthropic.

AISI’s assessments serve three functions: to preserve government’s ability to protect the realm in terms of cybersecurity, enabling a risk-based approach to AI integration in the economy; to provide public reassurance that AI technologies are understood and risk assessed; and to offer independent scrutiny of the performance of products. AISI’s ability to perform this triple function has rightly made it a critical element in the UK’s global AI influence. By cutting AISI out of the pre-deployment assessment process, Anthropic risks public trust at a time when it is most needed.

Why this decision?

The Financial Times alluded to pressure from the Trump Administration to adopt a more protectionist stance on AI products. This is a valid analysis: the US imposing export controls on Mythos and Fable and attempting to blacklist Anthropic’s participation in defense contracting, demonstrates that Anthropic is a particular target in a protectionist environment. The US AI Action Plan encourages a permissive innovation environment coupled with strong US protection. The trend of viewing AI as a national security and defence asset will only consolidate over time, and with it the level of export controls and sovereignty concerns associated with other defence technologies. The idea of robust assessments carried out by a foreign government simply do not align with an increasingly robust America First position, nor with the realities of national security asset classification, which among other restrictions also places nationality caveats on engineers participating in US defence and national security contracts. Anthropic is unlikely to be the last major US AI provider to downgrade its open relationship with AISI as a result.

However, the UK’s wait-and-see approach to AI regulation is vulnerable, depending on good will and cooperation to evaluate the most capable frontier models. Individual companies cannot be compelled to submit models to AISI for testing under current UK legislation, nor is there a legal framework in place to set the terms for what constitutes responsible and trustworthy AI (or even to prohibit the worst practises).

What are the solutions? When the UK government attempted to curb the power of communications service providers and compel them to provide lawful exceptional access to communications—even in threat to life cases—they continued to rely upon the good will and participation of US tech companies, despite operating in a well-legislated area. Ultimately, the partial solution to the challenge required many years of negotiation including the passing of additional UK legislation and the negotiation of data access treaties with the US government to facilitate access. Legislating a requirement for frontier AI models to subject themselves to AISI scrutiny presents a range of problems, not least compelling access to systems held in  US jurisdiction, and with a hostile administration to negotiate with.

Could companies that do not voluntarily submit to collaborate with AISI be prohibited from offering their services in the UK? Such plans have been mooted in the face of non-collaborative and potentially harmful US tech before, notably social media companies, but consistently been found as undemocratic, and unenforceable. For AI the U.K. cannot afford to fall behind its competitors in the AI race and so any ban might be seen as counter-productive. The tension between public safety and prosperity is ever present in all conversations about regulating and evaluating technology; the EU AI Act’s monitoring and enforcement challenges have laid bare the sheer scale of the challenge in this space.

As things stand, voluntary collaboration is an insecure method of securing compliance when the stakes are as high as they are in the development of frontier AI, especially given the unprecedented geopolitical backdrop  for UK and US tech companies . It is easier to argue and compel compliance with principles of responsibility, open and trustworthy scrutiny, and trustworthiness laid out in law and codes of practices, with an enforcement mechanism, however imperfect a method. That would provide a clear standardto show where an AI company falls short of the expectations for practicing and operating in the UK. The lesson of communications service providers looms large: the path to a satisfactory settlement with tech lies along establishing rules and litigating them in a long and challenging process; the negotiation of international treaties to unblock jurisdictional, defence and security, and data transfer restrictions; and concerted attempts to find common ground on international standards and approaches to responsible and trustworthy AI evaluation at the global level.   The sooner that process begins, the more effectively  evaluation bodies like AISI can provide assurance for national security, reinforce public trust, and demonstrate that private companies are subject to technical accountability for the safe and responsible operation of their models.