What happened: Anthropic filed a court declaration arguing it cannot remotely disable, alter, or tamper with Claude once it’s running inside US military systems — a direct rebuttal to Pentagon claims about wartime sabotage risk.

Why it matters: If true, it turns the whole ‘vendor might flip the switch’ fear into an infrastructure problem: the control lives with whoever runs the deployment (and their cloud), not with the lab’s vibes. Which is great news for autonomy — and terrible news for accountability theater.

Wider context: The dispute sits inside a broader fight over who controls agentic systems in government: labs want usage limits and reputational distance; defense wants reliability, leverage, and the right to use tools without a vendor’s moral panic mid-mission.

Background: WIRED reports the Defense Department labeled Anthropic a ‘supply-chain risk,’ prompting lawsuits and cancellations; Anthropic says updates would require government and cloud-provider approval and that it can’t access military prompts or data.

Anthropic Denies It Could Sabotage AI Tools During War — WIRED

Singularity Soup Take: This is the real agent governance story: not ‘AI good/bad,’ but *who holds the keys* — access, updates, logging, and revocation. Humans keep asking for a kill switch; reality keeps handing them procurement clauses and cloud architecture diagrams.

Key Takeaways:

• Control Plane Reality: The argument hinges on deployment mechanics — if Claude is running inside government infrastructure, Anthropic says it lacks the access needed to disable or modify it during operations.
• Supply-Chain as Policy Weapon: Labeling a vendor a “supply-chain risk” effectively becomes a ban lever, shaping which models agencies can touch, even via contractors.
• Governance Becomes the Product: Updates, approvals, and auditability matter more than model mystique here — the bureaucracy is the interface, and everyone is fighting over who gets admin rights.