Nick Trenkler
reuters reported this morning that senior anthropic technical staff are meeting with department of commerce officials in washington. they're trying to walk back a government order that forced anthropic to shut off its two most advanced models, fable 5 and mythos 5, to every foreign national in the world – including anthropic's own foreign employees
the government said it had found a way to "jailbreak" the models into helping find software vulnerabilities. anthropic says the flaws were minor and that any publicly available model can find them too
it's worth zooming out, because this is not an isolated incident. since january, the san francisco ai lab behind claude has been in an open structural fight with the trump administration over who decides how frontier ai gets used. anthropic is one of the few major labs that publishes hard usage policies – no autonomous weapons targeting, no mass surveillance – and the white house, the pentagon (rebranded "department of war"), and now the commerce department have spent six months trying to bend, bypass, or punish those policies. anthropic is suing in two federal courts. the timeline that led here:
- jan: reuters reports anthropic and the pentagon are at a standstill on a contract worth up to $200 m. the sticking point is anthropic's refusal to drop guardrails against autonomous weapons targeting and domestic surveillance. defense secretary pete hegseth had just released an ai strategy requiring all department of war ai contracts to permit "any lawful use," language designed to override company-specific restrictions
- feb: wsj reports claude was used in the maduro raid via anthropic's partnership with palantir – a deployment that arguably tested anthropic's own usage policies. the pentagon announces it's reviewing the $200 m contract. hegseth summons ceo dario amodei to a meeting characterized internally as "shit or get off the pot"
- mar: trump posts on truth social directing all federal agencies to cease using claude. hegseth formally designates anthropic a "supply chain risk," cutting it out of military systems and barring contractors from using its tech. anthropic sues the pentagon in two federal courts – northern district of california and dc circuit court of appeals – arguing the designation is unlawful and retaliatory. on mar 24, the nd cal court grants a preliminary injunction, finding the government's actions were "classic illegal first amendment retaliation" rather than national security protection
- apr: the dc circuit court of appeals denies anthropic's request to temporarily block the supply chain risk designation, but agrees to expedite the case, citing likely "irreparable harm"
- may: the pentagon announces classified-network ai deals with eight major tech companies – anthropic is the conspicuous exclusion
- jun: the export-control directive on fable 5 and mythos 5 lands. anthropic disables the models globally rather than partially restrict
which brings us to today's commerce department meeting.the underlying dispute is structural – whether a private lab or the state decides lawful military use of frontier ai – and anthropic has chosen to litigate rather than fold. one court has called the government's conduct first amendment retaliation; another has let the blacklist stand pending appeal. the pentagon has moved on to eight other vendors. and the export-control lever has now been added to the procurement one, on a thin public justification anthropic itself describes as discoverable using other public models
the question isn't whether the friction is real. it's how many more levers the administration is willing to pull before the courts catch up
