Pentagon sets end-of-week deadline for full military access to Anthropic's Claude
Trust is fraying between the Pentagon and Anthropic over access to the company's AI model, Claude. In a Tuesday meeting at the Pentagon, Defense Secretary Pete Hegseth gave Anthropic CEO Dario Amodei until 5 p.m. Friday to sign an agreement granting the military full access to the system.
Defense officials are weighing use of the Defense Production Act if talks stall. They also discussed designating Anthropic as a supply chain risk, which could sideline the company from future government work.
According to officials, the Department of Defense wants full control of Claude for lawful military use. Anthropic has pushed for guardrails that would prevent mass surveillance of Americans and block autonomous targeting decisions without a human in the loop.
A senior Pentagon official pushed back, saying, "This has nothing to do with mass surveillance and autonomous weapons being used. The Pentagon has only given out lawful orders." Hegseth also drew a comparison to Boeing aircraft: once the government buys a capability, the department decides how it's used.
Amodei has raised reliability concerns, noting Claude can hallucinate and should not be used to make final targeting calls without human judgment. Anthropic said it is continuing "good-faith conversations" to align usage with what the models can reliably and responsibly do.
Officials also said Grok, owned by xAI, is willing to operate in a classified setting, and other AI vendors are close. Anthropic, notably, was the first tech firm authorized to work on military classified networks and holds a $200 million Pentagon contract awarded last July.
Why this matters for government leaders
This isn't just a vendor dispute. It's a live test of how far agencies can go in asserting access rights over commercial AI, and how vendors can set safety limits without losing the contract.
Two policy anchors are in play: potential use of the Defense Production Act to compel access, and existing DoD rules on human involvement in weapon systems. For context, see the Defense Production Act Title III overview here and DoD Directive 3000.09 on autonomy in weapons systems here.
What agencies should do now
- Define non-negotiables: codify limits on surveillance of U.S. persons and on autonomous targeting; align with existing law and policy.
- Map access tiers: specify who can access models, weights, eval harnesses, logs, and deployment environments-and under what authorities.
- Write clear AI license terms: lawful use, audit rights, red-teaming and test data access, incident reporting, model versioning, and kill-switch expectations.
- Build human-in/on-the-loop enforcement: approval workflows, escalation paths, traceability, and immutable logging for sensitive use cases.
- Plan contingencies: identify alternate vendors, escrow critical artifacts, and include exit and re-competition triggers tied to milestone performance and trust.
- Assess supply chain risk: document vendor trust criteria, third-party assurance, and how a risk designation would affect missions and continuity.
Key dates and decision paths
- Friday, 5 p.m.: deadline for Anthropic to sign the access agreement.
- Possible outcomes: a signed deal with defined guardrails; invocation of the Defense Production Act; a supply chain risk designation; or a program pause and re-compete.
Bottom line
The department wants speed and control. The vendor wants clear safety limits. However this breaks, it will set a precedent for federal AI procurement and governance on classified networks. Get your negotiation playbooks, legal positions, and technical guardrails ready now.
For ongoing guidance on procurement, governance, and safe deployment, see AI for Government.
Your membership also unlocks:
