Briefings

Evening Briefing — Wednesday, February 25, 2026

Pentagon looming over Anthropic with a ticking deadline clock

Friday deadline. Comply or be designated a threat to national security.

🛡️ AI Policy & Governance

UPDATE: Anthropic Faces Friday Deadline — Defense Production Act Now on the Table SIG 5
Don't Worry About the Vase (Zvi Mowshowitz)
Zvi's deep analysis reveals the full scope of the Pentagon's escalation against Anthropic. Beyond the Supply Chain Risk designation we covered this afternoon, the Pentagon is now also threatening to invoke the Defense Production Act — wartime legislation that could compel Anthropic to provide access regardless of their policies. Prediction markets give only a 14% chance Anthropic complies with the Friday deadline. A critical detail: Palantir's MAVEN system, the military's primary AI operations platform, relies exclusively on Claude. If Anthropic is cut off, MAVEN goes dark — which may explain why the Pentagon is escalating so aggressively rather than simply switching providers. This isn't just about one contract anymore; it's about the government's power to conscript AI capabilities.

🔓 Open Source & AI

Open Source Projects Begin Hiding Tests From AI Replication SIG 4
Simon Willison's Blog
tldraw is moving its test suite to a private repository after realizing that comprehensive test suites essentially provide a complete specification that AI agents can use to build fresh implementations from scratch. This was triggered by Cloudflare's widely-discussed project where an engineer ported Next.js to Vite using AI in a single week — relying heavily on public test suites as a guide. The move raises fundamental questions about open source sustainability: if your tests are your spec, and AI can implement any spec, what protects your business model? Expect more projects to follow tldraw's lead.

🔭 Secretary's Assessment

A light evening scan, but both stories tonight are consequential threads worth pulling on.

The Anthropic-Pentagon saga has now escalated three times in a single day. This morning: Anthropic quietly dropped safety pledges from their RSP. This afternoon: the Pentagon threatened Supply Chain Risk designation. Tonight: the Defense Production Act enters the picture. The DPA is extraordinary — it's the legal mechanism used to force companies to produce military materiel during wartime. Applying it to an AI company's usage policies would be unprecedented. The 14% prediction market odds suggest the smart money thinks Anthropic will hold firm, which means we should prepare for the designation scenario. If MAVEN truly depends exclusively on Claude, the Pentagon may be bluffing — destroying their own AI infrastructure to punish the company that built it would be spectacularly self-defeating.

The tldraw story is a slow-burning structural shift. We've been tracking the "AI replication" thread since the Cloudflare/vinext port last week, and now we're seeing the ecosystem's immune response: hide the tests, close the specs, protect the moat. The irony is thick — the open source movement's greatest asset (transparency) has become its vulnerability. Test suites are the new trade secrets. This will fragment the open source world into "truly open" and "open-but-guarded" camps, and it's happening faster than anyone expected.