The full breakdown is at Skila AI.
What Actually Happened
In late February 2026, the Pentagon asked Anthropic to remove usage restrictions on Claude for mass surveillance and autonomous weapons decisions. Anthropic said no.
The next day, Trump ordered a 6-month phase-out of Anthropic across all U.S. government agencies.
Hours after that, OpenAI signed the deal Anthropic walked away from.
The Technical Split That Actually Matters
Anthropic: Claude's usage policy explicitly prohibits using it for systems that surveil populations or make autonomous lethal decisions. These are baked into the acceptable use policy — not contractual add-ons.
OpenAI: Three stated red lines around WMDs, indiscriminate civilian attacks, and undermining AI oversight. None explicitly block surveillance infrastructure or autonomous targeting chains.
MIT Technology Review described the gap as "OpenAI's compromise is exactly what Anthropic feared."
What the Pentagon Admitted
A senior defense official described the internal reaction when leadership realized how deeply Claude had been embedded in active workflows — summarizing intelligence reports, drafting analysis, processing classified documents.
They called it the "whoa moment." They granted a 6-month transition specifically because an immediate cutoff would have been too disruptive to active operations.
They banned Anthropic and simultaneously admitted the ban would hurt them. That's the detail most coverage missed.
xAI's Entry
Grok is being phased in alongside OpenAI as a government-approved alternative — giving Musk significant federal contract exposure while he's actively shaping federal AI policy through DOGE and advisory roles.
What This Means for Developers
If you're building on AI infrastructure and haven't modeled what happens if your vendor relationship ends suddenly, the Pentagon just demonstrated why you should.
Multi-vendor strategy is now an operational risk question, not a price negotiation.
Originally published on Skila AI with full analysis.
Top comments (0)