Anthropic has suspended global access to its newly released large language models shortly after making them public. The Claude Fable 5 and Mythos 5 models were announced and made available last week on Tuesday; three days later the United States government issued an export‑control directive, citing national‑security concerns and barring access for non‑US citizens. The restriction covers foreign nationals both inside and outside the United States, including Anthropic’s own non‑US employees.
According to Anthropic, the only practical way to comply with the directive was to stop providing access to the models worldwide to all users, including corporate partners that had already integrated the models into internal systems.
What triggered the restriction?
The government action followed vulnerability findings reported by Amazon researchers. Their work highlighted that the models’ advanced cyber capabilities — specifically the ability to find software vulnerabilities and generate functioning exploit code — could be misused for offensive purposes. After the report, government experts investigated the reported bypass and consulted with Anthropic executives.
Reporting in the Wall Street Journal and other outlets indicated that Andy Jassy, Amazon’s CEO, raised the issue with the government after his researchers identified the issue; Reuters cited David Sacks saying the export restriction was imposed after Anthropic CEO Dario Amodei refused to immediately patch the reported jailbreak or withdraw the model. The government said it hopes Anthropic will fix the issue so the restriction can be lifted.
Anthropic calls the restriction disproportionate
Anthropic rejected the government’s characterization. The company says the reported bypass — which the government presented only verbally — is not unique and reflects a capability level available in other commercial models. At the same time, Anthropic’s own measurements show Mythos 5 is materially stronger on cyber tasks than some competitors.
In a nearly 320‑page public system card, Anthropic describes the models and associated safety considerations. One test cited there tasked Mythos 5 with producing working exploit code targeting Firefox vulnerabilities; Mythos succeeded in 88.4 percent of cases, while Anthropic’s previous top model, Opus 4.8, succeeded in only 8.8 percent. The document also presents results from Anthropic’s ExploitBench benchmark.
On ExploitBench, Mythos 5 scored 10.75, Opus 4.8 scored 5.56, and OpenAI’s GPT‑5.5 scored 4.44. According to Anthropic’s comparison, Mythos’s score is more than twice that of GPT‑5.5. Observers note, however, that ExploitBench is an Anthropic benchmark and that cross‑vendor comparisons are complicated because different vendors generally do not use the same tests.
Model history and safety‑guard debates
The Mythos line traces back to Project Glasswing in April, when Anthropic provided about 50 trusted partners — including Amazon, Google and Microsoft — with access to a Claude Mythos preview. The company said at the time that the preview was not released publicly because of the model’s capabilities.
Last week’s public launch presented Mythos 5 as the stronger successor to the preview, while Fable 5 was positioned as a more constrained variant intended for broad public use. The system card, however, acknowledges that safety measures are not unbreakable. Testers at the UK‑based AI Security Institute reportedly found a one‑step bypass of Fable after a few hours of use, although more complex, agent‑style tasks did not defeat the model’s protections. Anthropic’s paid red‑team program logged roughly 100,000 attempts without finding a universal jailbreak that would broadly disable Fable 5’s cyber safety controls — a result Anthropic cites to argue the issue is narrow and the blanket recall disproportionate.
The debate therefore centers not only on raw model capability but also on how well developers can implement and maintain guards that prevent misuse while preserving legitimate security testing functions.
Political and industry context
The step follows months of increasingly fraught interactions between Anthropic and U.S. government bodies. Earlier this year reports said the U.S. military used Claude models in operational settings, and negotiations between the Pentagon and Anthropic over permitted uses reportedly became tense. In February, the Defense Department considered labeling Anthropic an supply‑chain risk and was reportedly close to terminating a contract potentially worth up to $200 million; Anthropic later sued and a judge temporarily blocked the designation.
The political context is evolving as well. President Donald Trump signed an executive order in early June creating a voluntary 30‑day pre‑release review process for companies to share new models with government reviewers — a different mechanism from the export‑control directive used in the Fable/Mythos case.
Not just a U.S. domestic issue: impact on Europe and companies
The government directive directly affects European users and companies, including Hungarian users, because it suspends access for non‑US citizens regardless of where the model was being used. The episode underscores the operational risk that reliance on U.S. frontier models can create: a single Washington decision can abruptly halt global access.
Europe currently lacks a widely available frontier model with a similar cyber‑capability profile to Mythos or Fable, meaning firms that depend on American models face strategic supply‑security questions. While Fable was publicly accessible for only three days before being suspended and thus unlikely to have been deeply embedded in many corporate workflows, a longer‑standing model being cut off could create serious operational disruption.
Next steps
U.S. officials said the restriction could be lifted if Anthropic remedies the vulnerability. Anthropic maintains the issue is narrow and disputes that it justifies the broad suspension. Ongoing technical evaluation and government‑company discussions will determine when and under what conditions Fable 5 might be reinstated for public use and Mythos‑level restrictions eased.
The episode highlights that frontier AI development is both a technological and geopolitical issue: national decisions can quickly translate into global access changes, and model choice now carries not only technical but also supply‑security implications.
