In this issue: Anthropic rejects Pentagon → Claude's consumer growth → Proton Mail privacy incident → AI engineering anti-patterns → Lenny's insights → BlackRock credit fund limits redemptions → References
Anthropic's Two-Front War: Ethics vs the Pentagon
The biggest Anthropic news this week wasn't technology — it was politics.
The Pentagon officially designated Anthropic a "supply chain risk" after the two sides couldn't agree on AI model control — the military wanted unrestricted use of Claude, including for autonomous weapons and mass domestic surveillance. Anthropic refused. The $200M contract collapsed.
AI models are rapidly commoditizing. Anthropic is positioning itself as the "ethical and trustworthy AI provider" — and that has real market value for both consumers and enterprises. — Bruce Schneier & Nathan Sanders
Ironically, Claude's consumer growth actually accelerated. TechCrunch reports Claude App new installs have surpassed ChatGPT, with Anthropic's annual revenue reaching $19B ARR.
Meanwhile, OpenAI's "compromise" placed Anthropic in a more complex position: MIT Technology Review reported OpenAI accepted some Pentagon terms — precisely the precedent Anthropic feared.
Proton Mail Privacy Incident: Encryption Doesn't Mean Anonymity
A widely discussed case: Proton Mail assisted the FBI in unmasking an anonymous "Stop Cop City" protester.
The core logic: Proton Mail encrypted email content, but IP addresses and backup recovery emails were not in scope. Under legal demands, Proton provided this "metadata."
This is a reminder: threat models matter. Encrypted content does not equal complete anonymity.
AI Engineering Anti-Patterns: Simon Willison's Warnings
Simon Willison organized the most common developer mistakes in his "Agentic Engineering Patterns" guide:
- Inflicting unreviewed code on collaborators: Don't submit PRs you haven't read yourself
- Over-trusting AI test results: Agents can execute code, but that doesn't mean they understand what the test is testing
- No Prompt Injection protection: Any agent receiving user input must be defended
- Sharing cache keys across workflows: The Cline case lesson (see W04)
Lenny's Newsletter: Multiple PM and Founder Insights
This week Lenny interviewed multiple product and design thought leaders:
- Aaron Levie (Box CEO): "Every Agent Needs a Box" — enterprise AI agents need content management as infrastructure
- Jenny Wen (Claude head of design): The design process is dead — Claude's designers no longer do wireframes; they prototype directly
- Multiple pieces on debugging a "team that isn't working" in the AI era
BlackRock $26B Private Credit Fund Limits Redemptions
A notable financial market signal this week: BlackRock limited redemptions on its $26B private credit fund — client withdrawal requests surged to 9.3% of NAV.
PIMCO simultaneously warned that private debt faces a "full-blown default cycle." Fed Cleveland President said rates could stay unchanged for a long time.
References
Anthropic vs Pentagon
- Anthropic and the Pentagon — Simon Willison (03/07)
- Claude's consumer growth surge continues after Pentagon deal debacle — TechCrunch (03/07)
- When AI Companies Go to War, Safety Gets Left Behind — Wired (03/07)
- OpenAI's "compromise" with the Pentagon is what Anthropic feared — MIT Technology Review (03/03)
Proton Mail
- Proton Mail Helped FBI Unmask Anonymous 'Stop Cop City' Protester — Hacker News (03/06)
AI Engineering Anti-Patterns
- Anti-patterns: things to avoid — Simon Willison (03/05)
- Agentic manual testing — Simon Willison (03/06)
Lenny's Newsletter
- Every Agent Needs a Box — Aaron Levie, Box — Latent Space (03/05)
- The design process is dead — Jenny Wen — Lenny's Newsletter (03/01)
Financial Markets
- BlackRock $26B Private Credit Fund Limits Withdrawals — Bloomberg (03/07)
- Private Debt Should Face 'Full-Blown Default Cycle,' Pimco Says — Bloomberg (03/07)