Two weeks ago, Anthropic's two most powerful models went dark. Mythos 5 and Fable 5 were both pulled after security researchers found a method to bypass Fable 5's safety guardrails — a vulnerability that could unlock Mythos 5's advanced cyber capabilities. On June 26, Commerce Secretary Howard Lutnick sent a letter to Anthropic's chief compute officer Tom Brown: Mythos 5 is back, but only for a select group.
More than 100 US companies and federal agencies made the "trusted partners" list. Lutnick's letter states that "appropriate safeguards are in place" — the key phrase that authorized access. The clearance extends to non-American employees working within those organizations, a partial reversal of the original ban's blanket restriction on any foreign national accessing the model.
Fable 5 remains completely off limits. Lutnick's letter doesn't mention it. The consumer-facing model is harder to confine to controlled environments, and it was the original entry point for the exploit researchers found. There's no sign that public access is returning anytime soon.
What's genuinely new here isn't the restriction — it's the mechanism. Governments have long controlled physical exports: chips, data center licenses, satellite hardware. Applying that same logic to neural network weights is different territory. The Commerce Department is now acting as a gatekeeper deciding who can run a specific AI model. The closest analogy is ITAR, the export control regime for military technology, applied for the first time to software at this scale.
For the 100+ organizations on the list, Mythos 5 is operational again. For everyone outside it — competitors, international researchers, startups that didn't qualify — the ban continues. Anthropic's challenge now is keeping the partner list confidential; publishing it would make those organizations targets. The question the episode raises but doesn't answer: what happens when the next generation ships with capabilities that make Mythos 5 look modest?
