The first major AI 'blackout' in the U.S.: Anthropic's Fable 5 returns after two weeks of government veto

By Momentum IA · June 28, 2026.

For more than two weeks, Anthropic's most powerful AI model —Fable 5— was formally banned for users outside the United States by order of the Trump administration. According to Axios, citing internal company sources, the administration claimed to have identified a method to circumvent ('jailbreak') one of Fable 5's safeguards, specifically the one that prevents the model from being used to discover software vulnerabilities. Commerce Secretary Howard Lutnick confirmed in a letter published by Semafor that negotiations between the government and Anthropic have yielded 'significant' results, and that the model should be available abroad again within days.

The specific facts deserve attention. The ban took effect on June 12, after months of tension between U.S. officials and Anthropic over access to the technology, regulation and security mechanisms. The second model involved, Mythos 5 —described as Anthropic's cybersecurity-specialized model—, was not banned but was restricted to authorized individuals under strict regulations. Anthropic maintains that Mythos 5 now operates under 'strong safeguards' that would make it unfeasible to use for cyberattacks or biological weapons development, the two central concerns expressed by Washington.

Anthropic was emphatic in rejecting the logic of the ban: it characterized the decision as based solely on 'verbal evidence of a potential narrow and non-universal jailbreak,' and argued that the existence of a theoretical and limited vulnerability does not justify withdrawing a commercial model on which hundreds of millions of people depend. It is an understandable position and, in a sense, technically reasonable: no complex system is one hundred percent impenetrable, and the threshold the government applies could paralyze any technology if taken to the extreme. That said, the government has a legitimate point: a model with Fable 5's ability to identify software flaws in the wrong hands represents a systemic risk that is qualitatively different from that of earlier tools.

The most relevant outcome is not the model's return, but what both parties have agreed for the future. Anthropic has committed to working with the government on protocols, standards and release processes. Both Anthropic and OpenAI are pushing to codify Trump's executive order on 'voluntary review of new AI models,' which in practice would mean giving regulators formal power to block developments before deployment. The financial context matters: both companies have IPOs planned before the end of 2026, and neither can afford an open regulatory battle that scares off investors or blocks international markets. Technological diplomacy is also stock-valuation diplomacy.

Our read: this episode is a turning point in the relationship between democratic governments and frontier AI labs. Until now, regulation has been mostly reactive and symbolic —general legal frameworks, voluntary guidelines, statements of intent—. Here we see something qualitatively different: an administration flexing executive muscle to shut down a specific product, deployed at global scale, on national security grounds. The fact that the ban lasted only two weeks and was resolved through negotiation suggests the system can work: government and company found a common language. But the precedent is set. The next lab to release a model with dual-use capabilities —cybersecurity, synthetic biology, military reconnaissance— now knows it could face an executive ban overnight.

In the short term, this introduces friction and uncertainty for developers and enterprise users building on these platforms. A company that integrates Fable 5 into its production systems cannot afford to have its most critical tool disappear without warning for two weeks; that pushes them to diversify providers and to demand contractual guarantees that the market does not yet know how to articulate. Reliance on single frontier models suddenly becomes a real operational risk.

In the long term, paradoxically, this kind of regulatory friction can be healthy. The alternative —models with mass-destruction capabilities released without any oversight— is worse for everyone, including the industry. The agreement reached, with joint standards and pre-launch vetting processes, is exactly the kind of trust framework that the most powerful models need to keep being deployed without provoking a more drastic and indiscriminate legislative backlash. If Anthropic and OpenAI manage to make that framework rigorous but technically informed —rather than an arbitrary political ban—, they will have done the sector as a whole a service.