Open Letter Opposing the Suspension of Australian and US Allies Access to Frontier AI Models

Summary

On 12 June 2026, the United States government issued an export control directive that suspends all access to Anthropic’s most capable AI models — Fable 5 and Mythos 5 — for every foreign national, whether inside or outside the United States. To comply, Anthropic disabled both models for all customers worldwide. Australia was given no notice, no consultation, and no differentiated treatment. We learned of a decision affecting our national AI capability at the same moment as the general public.

We are writing to ask the Australian Government to treat this as the strategic event it is, and to make urgent, direct representations to the United States for an explicit allied carve-out. The window to do so is measured in days, and the cost of inaction compounds.

This is both an incident and a precedent. The incident is serious in itself: these are the most advanced general-purpose models available globally, and the tier of capability around which Australia’s own National AI Plan and its bilateral agreement with Anthropic were oriented. The precedent is graver still: this is the first live demonstration that frontier AI access can be switched off for allies unilaterally, without prior notice or consultation. The capability to do this again is now established, and Australia has no say in when it is next used.

Why these models matter

This is not a dispute over a niche commercial product. The class of models now withdrawn from allied access has, on the public record, become an instrument of state power.

• US forces used Anthropic’s Claude, via its partnership with Palantir, during the January 2026 operation that captured former Venezuelan President Nicolás Maduro — the first publicly reported use of a commercial AI developer’s model in a classified US military operation.¹
• In the joint US–Israel strikes on Iran beginning in late February 2026, Claude was used alongside Palantir’s Maven system for real-time target identification and prioritisation; reporting describes the model as embedded in US military intelligence and targeting workflows.²

Whatever one’s view of those operations, they establish the point beyond argument: frontier models of this class are now geopolitically central. A government’s ability to access them — or to have that access revoked by another government — is a matter of national capability, not consumer convenience.

These specific models are the leading edge of that class. Anthropic released Fable 5 on 9 June 2026 as the first publicly available model of its most capable (“Mythos-class”) tier, positioning it for the most demanding knowledge and engineering work; senior engineers described it as the largest single capability step in recent model generations.³ It is precisely this frontier tier to which Australia has now lost access.

What happened

On 12 June 2026, Anthropic received a directive from the US government at 5:21pm (ET) requiring it to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including Anthropic’s own foreign-national employees.⁴ The directive was issued by the US Department of Commerce, via a letter from Commerce Secretary Howard Lutnick to Anthropic’s CEO, under export-control authority — making a licence a precondition of any export, re-export, or domestic transfer of the models.⁵

The stated basis was a national security concern relating to a method of “jailbreaking” Fable 5. On the public record, the triggering technique amounts to asking the model to read a codebase and identify software flaws — a capability Anthropic states is widely available from other deployed models, including OpenAI’s GPT-5.5, and used daily by the defenders who keep systems secure.⁴ Anthropic is complying with the legal directive while publicly disputing its basis, and has stated it believes the matter is a misunderstanding it is working to resolve.⁴

The effect on allies was indiscriminate. The control applies to all foreign persons.⁵ An Australian national — a citizen of a Five Eyes partner and an AUKUS member — sits in precisely the same category as the adversary states the export-control regime exists to contain.

Why this is a strategic event, not a commercial disruption

It would be a mistake to file this as a vendor outage affecting a single product. Three features make it a matter of national strategic consequence.

A formal bilateral commitment has been disrupted. On 31 March 2026, the Australian Government signed a Memorandum of Understanding with Anthropic to cooperate on AI safety research and to support the goals of the National AI Plan, formalised when the Prime Minister met Anthropic’s CEO in Canberra.⁶ Central to that agreement is a commitment to work with Australia’s AI Safety Institute — sharing findings on emerging capabilities and risks, participating in joint safety and security evaluations, and collaborating with Australian research institutions.⁶ A US directive has now made that cooperation impossible for exactly the frontier models where it matters most, and it was reached without consultation of the Australian Government.

The precedent outlasts the incident. Whatever happens to Fable 5 in the coming weeks, the demonstrated fact is that allied access to frontier AI can be terminated by another government unilaterally and without warning. For a nation building its AI strategy on access to overseas frontier models, that is a structural vulnerability, not a one-off inconvenience.

The allied safety relationship was already operational and has now been cut. In the weeks before launch, Anthropic states that the UK AI Safety Institute was among the bodies that red-teamed Fable’s safeguards for thousands of hours.⁴ Allied safety institutions were inside the testing process. A blanket ban removes precisely the allied researchers whose evaluation work makes these models safer — including the joint evaluations Australia’s own AISI is committed by MOU to perform.⁶

Why an indiscriminate ban harms the security interest it invokes

The directive does not remove frontier capability from the world. By Anthropic’s account, the capability that triggered it is already available from other deployed models.⁴ What the directive removes is the US-allied supply. The demand does not disappear; it relocates — and for an allied nation, there is no neutral destination to relocate to. The only competitive non-US source of frontier large language models is China.

A blanket lockout of allies therefore does not produce caution. It produces an immediate, forced channelling of allied demand, investment, and dependence toward Chinese AI infrastructure. This is not a gradual long-term drift; it is a short-term structural shift, and it hands Beijing the single outcome the export-control regime exists to prevent — the democratic world building on, and becoming dependent upon, an adversary’s frontier technology. A measure justified on national security grounds, applied indiscriminately to allies, is self-defeating on its own terms.

The constructive alternative is already articulated in US policy thinking — and the present directive is its precise inversion. In Policy on the AI Exponential (June 2026), Anthropic’s CEO argues that democracies should form a coalition built on shared values, and grow it by making membership more and more attractive and remaining outside it less and less attractive — starting with aligned democracies and progressively drawing in less-aligned nations prepared to meet the coalition’s standards, until, ideally, the entire world joins.⁷ Shared access to frontier models is the central attractive force in that design: it is the benefit that makes alignment with democratic standards worth choosing. The doctrine is explicit that the supply chain should be secured by sharing it within the coalition and denying it to those outside it.⁷

An indiscriminate ban on allied access does the literal opposite. It denies the supply inside the coalition — to Five Eyes and AUKUS partners — while doing nothing to deny it outside, since the same capability remains available elsewhere. In doing so it sabotages the very mechanism by which the democratic world is meant to draw others in: it makes coalition membership demonstrably less attractive — an ally can be cut off without notice — at the same moment it pushes undecided and developing nations toward the only available alternative, which is China’s. A measure justified on national security grounds thus undermines, on its own stated terms, one of the most powerful instruments the United States has for drawing the rest of the world toward democratic standards. An allied carve-out is therefore not special pleading; it asks the United States to act consistently with the strategy its own leading safety-focused developer has publicly set out.

Why this cuts against Australia’s stated national policy

In December 2025, the Albanese Government released the National AI Plan, built on three goals: capturing the economic opportunities of AI, spreading its benefits broadly, and keeping Australians safe.⁸ A foreign-imposed, indiscriminate cut-off of frontier models is incompatible with every limb of that plan:

• It defeats the “capture the opportunities” goal by removing the most capable tools from Australian businesses, researchers, clinicians, and the technology sector overnight.
• It undermines the “keep Australians safe” goal by severing the joint safety evaluations the AISI was funded and committed to perform.
• It contradicts the ambition to be a competitive destination for AI investment by demonstrating that Australian access to frontier capability is contingent on decisions Australia has no part in.

The plan committed A$29.9 million to establish the AI Safety Institute in early 2026 as an expert body to test high-risk models and advise on safe deployment.⁸⁹ That institution is being undercut before it is fully operational: the Government has invested public money in an evaluation capability that this directive prevents from doing its core work on the models that most warrant it. Minister Ayres has framed AI access as a matter of economic resilience and competitiveness, arguing Australia must “lean in” and secure a stake in the technology as other nations compete for advantage.¹⁰ This episode is the concrete test of that position. Securing a stake means ensuring that stake cannot be revoked without consultation.

What we ask

We do not ask the Australian Government to oppose export controls in principle. Targeted controls that deny frontier capability to adversaries are legitimate and necessary. We ask the Government to insist that allies are not treated as adversaries. Specifically, we ask the Government to make urgent representations to the United States for:

1. An explicit allied carve-out from the directive for AUKUS and Five Eyes partners, restoring access for Australian nationals and Australian-based entities — and, importantly, structured so that it can be extended to other democratically aligned nations as appropriate, rather than leaving the rest of the democratic world locked out and driven toward adversary providers.

2. A defined and short restoration timeline for allied access, rather than an open-ended suspension, with transparency about the technical findings that would resolve the underlying concern.

3. Activation of the existing Anthropic–Australia MOU channel, so that the Australian AI Safety Institute participates directly in the joint safety and security evaluations of these models — converting the present lockout into exactly the kind of allied red-teaming both governments have already committed to.⁶

4. A standing allied-consultation mechanism so that any future export-control action affecting allied access to frontier AI is preceded by notice and consultation, not delivered as a fait accompli.

The negotiating window appears to be open now and short: an administration official, as reported by Axios, indicated the models would remain locked down only until the US national security apparatus is “hardened,” which the official suggested could occur within a few weeks.⁵ Allied representations made in the coming days can shape the terms of restoration — including whether an allied carve-out is built into the resolution and the precedent it sets. Once this is treated as “resolved” without Australian input, the opportunity to establish that precedent is gone, and the switching costs already accruing across the Australian economy become harder to reverse.

This is the most strategically consequential class of technology our governments have had to contend with. We urge the Government to act with corresponding urgency.

References