The Pentagon-Anthropic story has escalated from a policy dispute to something much darker. This morning we covered Anthropic quietly dropping safety pledges from their RSP. This afternoon, we learn the Pentagon is wielding the "supply chain risk" designation — a weapon designed for foreign adversaries — against a domestic AI company for having too many safety guardrails. The irony is suffocating: Anthropic is being punished by the government for the very principles that made it the "responsible" lab, while simultaneously abandoning those principles under commercial pressure. They're getting squeezed from both sides.

The deanonymization paper should concern anyone who's ever posted under a pseudonym — which is most of the internet. The technique isn't theoretical; it works at scale across real platforms. Combined with the Pentagon's push for unrestricted AI access, the picture is clear: the tools for mass surveillance aren't coming. They're here. The question is who deploys them and under what constraints.

Anthropic entering the always-on agent space with Claude Code Remote Control and scheduled Cowork tasks is strategically interesting. They're building toward persistent agency but with a safety net — tasks only run while your computer is awake. It's a measured approach compared to OpenClaw's "always-on" philosophy, and the Tachyon blog post about sandbox limitations suggests the security community thinks nobody has the containment problem solved yet. The agent security question may become 2026's defining infrastructure challenge.

ByteDance open-sourcing deer-flow continues the pattern we've been tracking: Chinese labs are releasing agent infrastructure at speed, creating a parallel ecosystem. When the two largest agent framework ecosystems (US and China) evolve independently, interoperability becomes a geopolitical question, not just a technical one.