On 12 June 2026, Anthropic disabled access to its Mythos 5 and Fable 5 models for its customers globally. This was reportedly in response to a US government export control directive requiring Anthropic to limit access to these models to US nationals only (including within Anthropic's workforce), citing national security concerns. As of 15 June, the nature of the national security concern is not clear and it is possible that the situation will be resolved quickly.
Mythos 5 is Anthropic's most advanced model, which demonstrates significantly improved ability for cyber security use cases compared with previous state-of-the-art models. Anthropic has made much of its coding capability and the potential cyber risks that the model presents, and initially offered it only to a small group of trusted customers. Fable 5, launched just last week, is a general-purpose offering, built on the same model as Mythos, but with additional technical restrictions to reduce the risk of its being used for dangerous or illegal purposes (including offensive cyber security applications).
This article considers the legal impact of this decision, how it fits into the UK and EU regulatory context, and some considerations for lawyers working on AI contracts.
History suggests that export controls on software are difficult to sustain in the long-term. We consider the analogy with the "Crypto Wars" of the 1990s, and whether regulation can hold back the tide of AI.
Brussels is likely to see this as a vindication of its emerging "AI sovereignty" agenda. The proposed EU Cloud and AI Development Act (published earlier this month) seeks to ensure that the EU's access to strategically important AI is not at the discretion of foreign states.
Lawyers working on AI contracts need to bear in mind the shifting regulatory environment when drafting and negotiating. In particular, force majeure, export control, compliance with law, and termination/suspension clauses need fresh consideration, especially where a deal involves access to frontier models.
"AI Wars"? Lessons from restricting cryptography
Those who have worked in the field for long enough may have a sense of déjà vu. During the "Crypto Wars" of the 1990s, the US Government (along with other governments, including in the UK) attempted to use export control regulations to restrict the proliferation of strong encryption. In practice, this policy did not prevent functionally unbreakable cryptography from becoming mainstream: internet connections and data storage are now routinely encrypted using algorithms that would have been illegal to export from the US in the 1990s.
The key reason for this failure was that cryptography became commoditised and freely available around the world (PGP being the most famous example, which led to its author being investigated for export of "munitions"). At that point, it became clear that the restrictions were putting US technology companies at a disadvantage without yielding any real national security benefit.
Could attempts to use export controls to limit access to powerful AI models meet the same fate? There are two obvious differences from the cryptography example, at least for the moment. The first is that the proprietary frontier models are provided as a service over API rather than distributed as software, meaning that the provider can control access (just as Anthropic has done in this case). The second is that frontier models usually require data-centre scale hardware to run, meaning that even if bad actors could obtain the model and weights, they would need the resources of a large corporation or a nation state to use it.
But these are not perfect defences. For one thing, nation states (or groups backed by nation states) are one of the key threat actors that authorities have in mind. More generally, open-source models such as Deepseek and Qwen are continually climbing the benchmarks, often outperforming larger, closed models with much lower hardware requirements (albeit usually several months later). If this trend continues, it may soon be possible to have Mythos-like capability in an open model, running on local hardware and at a cost that is accessible to a well-resourced criminal.
AI sovereignty and the UK and EU regulatory agendas
Digital sovereignty, and AI sovereignty in particular, has been a major talking point of the last few years for both policymakers and businesses. The US Government directive, and Anthropic's response to it, provide a clear illustration of why. AI safety and alignment researchers have long predicted that there would come a point where frontier AI models would become so competent in the field of cyber security that governments would feel compelled to restrict – or even monopolise – access to them. Countries that lack domestic technical capability risk being left out.
The UK has announced that this is a policy priority and recently launched a sovereign AI fund which will invest in UK-based start-ups and scale-ups. Earlier this month, the European Commission released its proposal for a Cloud and AI Development Act (CADA), which aims to "reduce critical external dependencies by strengthening homegrown cloud and AI capabilities and infrastructure". This follows hot on the heels of the Chips Act and the recently proposed Chips Act 2.0, which aim to provide supply-side and demand-side support for EU-based semiconductor manufacturing.
CADA contains measures to support AI research (particularly in the field of cyber security) and expand data centre capacity within the EU, with funding expected to come from existing EU programmes as well as from member state governments and the private sector. It also contains measures to push public-sector (and potentially even private-sector) procurement decisions toward EU-based cloud providers.
While the US Government's decision may vindicate those policymakers concerned about AI sovereignty, it remains to be seen whether the EU will be able to legislate and subsidise a thriving domestic cloud and AI ecosystem into existence this late in the game.
Contracting for AI amid regulatory uncertainty
Regardless of how the situation with Mythos and Fable is resolved, we doubt this will be the last example of a regulatory intervention affecting the cross-border provision of AI services. Lawyers working on AI deals will need to bear this in mind, and consider whether the contract is properly allocating the risks of regulatory change.
Most vendors' standard terms place all of the risk of such an event with the customer. The Cloud Legal Project (Queen Mary University of London) recently released a paper analysing standard terms from the key AI vendors. Unsurprisingly, these terms are very protective of the vendor in this scenario:
- Where standard terms address export control, they are frequently non-specific and the customer promises to comply with export control laws (which underplays the role of the provider in enabling compliance).
- Some providers set out a list of supported or unsupported regions which the provider is free to change during the contract term.
- Many providers reserve the right to terminate the contract for convenience and to suspend the services where required by law.
- Government orders are often treated as force majeure events, relieving the provider of its performance obligations while usually leaving the customer's payment obligations intact.
Our experience is that these contracts can be negotiated where a deal is sufficiently large or strategic. In this case, customers may wish to consider whether a more balanced position could be reached, for example by:
- Explicitly addressing regulatory changes: ensuring that they are not treated as force majeure events and that instead the risk is commercially allocated. While it clearly wouldn't be reasonable to expect the provider to continue providing a service in breach of the law, it would be reasonable to seek proper notice, obligations to try to overcome the disruption, and swift resumption of service when possible.
- Including commercial remedies: where access to cutting-edge models is a core part of the deal, including stronger rights such as termination without penalty if those models become unavailable, or potentially a fee reduction (akin to service credits) if the service falls back to less capable models.
- Sensible sharing of compliance obligations: resisting boilerplate "compliance with law" wording. Where the parties know that the service may be subject to export controls, customers could seek a more workable regime where the service provider is responsible for identifying whether their product is subject to export control restrictions and obtaining the relevant export permits, and the customer promises to comply with the end user restrictions in the permit.
For practitioners, the Mythos and Fable episode is a prompt to have a different conversation with clients who rely on frontier AI. The instinct is to reach for the contract, but the more valuable advice often comes earlier. Clients should be asked where frontier models actually sit in their business, how exposed they are if access is restricted or degraded overnight, and whether they have a realistic fallback. Many will not have thought about it in those terms until prompted.
The drafting points above then follow naturally from that conversation. But practitioners add the most value by treating this as a business-continuity issue first and a drafting issue second. Advising a client to design for substitutability, and to avoid building a critical process around a single model they may lose access to, will often matter more than any clause. Our view is that the risk of unpredictable national security or other regulatory interventions over frontier models is here to stay, and the firms that serve clients well will be the ones raising the question before the next directive lands, not after.
Copyright & permissions
