Models At The Border

The strange part is not that Washington is anxious about frontier AI. It is that the action treats model access itself as an export, not only the chips, tooling or cloud infrastructure used to train and run it. Axios reported that Commerce Secretary Howard Lutnick sent a letter to Anthropic CEO Dario Amodei saying Mythos 5 and Fable 5 would need export controls for any location outside the United States and for foreign persons inside the country. Axios also reported that a license would be required for export, re-export or domestic transfer of the models. This is a step up from previous AI and technology export controls. U.S. AI export policy has primarily focused on chips and tools rather than on limiting direct access to AI systems.

The complication is that Anthropic made the government’s case to regulate export easier. In its June 9 launch post, Anthropic described its Fable as a Mythos-class model that can work autonomously for longer than prior Claude models and that Mythos 5 has “the strongest cybersecurity capabilities of any model in the world.” Complicating things is that the company said its available-for-everyone Fable and highly restricted Mythos shared the same underlying model, with only safeguards marking the difference.

That design, which depends on the technology for policing itself, created a policy trap. If the same engine can run with softer brakes in trusted channels, regulators may ask whether the brakes can be stripped, bypassed or misused. Anthropic’s response was that it had tested Fable heavily with U.S. and U.K. authorities, private third parties and internal teams. It said no tester had found a universal jailbreak.

That argument is procedural. Anthropic is not denying the risk. It is just challenging the threshold for intervention, arguing that “possible” should not be treated as “proven.” Anthropic wrote that the disclosed jailbreaks were either benign or minor, with no Mythos-specific lift. It warned that if a narrow jailbreak could thwart a commercial model, the same standard could be applied to all of its model releases. So who should we trust in this instance, the company building and selling a model, or a government with different incentives to restrict or enable access?

Customers See The Cost Before Regulators Do

The access battle comes at a critical moment for Anthropic. Earlier in June, the company confidentially submitted a draft S-1 for a proposed IPO, giving it the option to go public after SEC review. It was also dealing with customer friction around Fable’s data rules. Anthropic required 30-day retention for all traffic on Mythos-class models, a change it said was needed to research and mitigate jailbreaks. Reporting from The Verge and Reuters said that Microsoft had limited employee use of Claude Fable 5 over data retention concerns.

With governmental restrictions and regulations, usage of AI models will get even messier. For banks, law firms, software companies and research labs, risks of access now sit next to latency, price, accuracy and security in reviews. Backlash over Anthropic’s plan to degrade responses for prompts linked to high-risk responses has been swift with some calling it appalling and hostile to users.

The other irony is that Anthropic has long argued for tougher AI oversight. Reuters noted that the company had recently called for greater U.S. oversight, including the power to block models with unacceptable risks. Yet Anthropic said this recent US export action did not meet fair, fact-based standards.

This tension between the desire to increase access to powerful models that can generate significant revenues for AI model companies and benefits for users on the one hand, and regulators who want to restrict access for safety and sovereignty issues on the other is becoming the main story. AI companies want to continue to release increasingly powerful models at faster, competitive rates, governments want control over systems that could aid cyberattacks, biological research or military planning, and customers want access to accelerate their transformation efforts that won’t vanish overnight.

Fable 5 may return. Anthropic says it is working to find ways to restore access. The harder repair will be user and government confidence. Frontier AI has crossed from product category into strategic asset, closer to aerospace, cryptography and nuclear-adjacent computation than ordinary software. And the fight is not just over Fable.

The shadow of government control now looms over all frontier AI models. Who will emerge as the winner: multibillion-dollar companies with global ambitions, users who want to transform their organizations and achieve greater productivity or governments with the power to decide where the technology can go, who can use it and when access must stop?