How the Mythos Shutdown Happened: Amazon CEO Jassy, SK Telecom, and 90 Minutes
The Mythos export ban that pulled two Anthropic models from every market outside the United States came down in two steps, and the specific sequence is now clearer.
Anthropic had extended Mythos access to approximately 150 vetted companies and government organisations under its limited partner program. The first trigger: one of those partners was a South Korean telecom that U.S. officials identified as suspected of having ties to China. The company — widely reported as SK Telecom, which has denied the connection — was enough to prompt a review.
The second trigger: Amazon CEO Andy Jassy personally alerted the Trump administration that Amazon’s own researchers had found a way around Fable 5’s safeguards. Anthropic disputed the characterisation. The company said the issue was narrow and already patched, not a wholesale defeat of the model’s safety measures. The administration treated it as a “jailbreak” regardless.
The result was a Commerce Department export control directive. Anthropic had roughly 90 minutes from notification to full shutdown of both Fable 5 and Mythos 5 — globally, for all users, including domestic ones who were not subject to the export restriction.
The Speed Problem
The 90-minute window is notable. Export controls typically operate at the scale of supply chains and licensing regimes — months of review, formal notifications, regulatory processes. Applying that framework to software deployable via API required Anthropic to make a company-wide operational decision faster than most incident response playbooks are designed to handle.
The result was an abrupt cutoff affecting approximately 200 organisations that had relied on the models, including research institutions, government agencies, and enterprise customers in allied countries. G7 partners subsequently proposed a “trusted-partner” access framework to avoid the same blunt instrument being applied to future model generations.
The PGP Comparison
The episode draws obvious parallels to the PGP Crypto Wars of the early 1990s, when the U.S. Customs Service opened a criminal investigation against PGP creator Phil Zimmermann for alleged arms export violations. Zimmermann ultimately published PGP’s source code as a printed book, exploiting the First Amendment protection for printed material in a way that made the export prohibition meaningless. The investigation closed in 1996. End-to-end encryption proliferated globally regardless.
The AI case is materially different. PGP was distributable software that, once released, could not be recalled. Frontier AI models are predominantly accessed through APIs under vendor control — Anthropic can actually enforce an access cutoff in a way Phil Zimmermann’s critics could not. Whether that control holds over time, as open-weight models at comparable capability levels reach the market, is the open question.
DeepSeek V4 Pro, released under MIT license with 80.6% SWE-Bench Verified, is already freely downloadable. Kimi K2.7 Code ships open-source. The enforcement window for capability-based AI export controls may be narrowing faster than any regulatory framework can move.
What Comes Next
The White House reportedly opened a review process to restore Mythos access to vetted organisations, with G7 allies pushing for a structured access pathway. Markets were pricing roughly 67% odds of partial restoration by July 1 as of last week.
Anthropic, for its part, has disputed both the jailbreak framing and the characterisation of Fable 5’s capabilities as uniquely dangerous. The company’s public position is that its safety messaging — which deliberately emphasised Mythos’s power to justify restricted access — may have contributed to the political conditions that produced the ban.