After Anthropic CEO Dario Amodei refused to drop two contractual red lines — no domestic mass surveillance, no fully autonomous lethal targeting — the administration moved against the company directly.

Anthropic called the action unprecedented and legally unsound and said it would challenge it. The same day, OpenAI announced its own agreement to put models on the Department of War's classified networks.

Perplexity introduced Computer, a cloud computer-user agent that routes each subtask to whichever of 19 models fits best — Gemini, Claude, GPT, Grok, Veo and more — and spins up sub-agents that it says can run workflows for hours or months, each in an isolated environment with a real filesystem, browser and tool integrations.

It ships only on the $200/month Perplexity Max tier. Ars Technica frames it as a contained answer to the viral OpenClaw local-agent tool; TechCrunch notes Perplexity canceled the press demo hours beforehand after flaws were found, and that the economics of routing across many models at a flat subscription are unproven.

Composio released Agent Orchestrator, an open-source orchestration layer that is itself an agent: it reads a repo and backlog, decomposes features into parallel tasks, spawns coding agents in isolated git worktrees and tmux sessions, then auto-injects CI failures and review comments back into the right session.

It is agent-agnostic, with adapters for 20+ harnesses including Claude Code, Codex, Aider, OpenCode, Cursor and Copilot. Composio says the 40,000-line TypeScript project with 17 plugins and 3,288 tests was built in eight days mostly by the agents it manages, with every commit carrying a git trailer naming the model that wrote it.

AMD and Meta announced a multi-year, multi-generation partnership to deploy up to 6 gigawatts of AMD Instinct GPUs, starting with the MI450 inference-optimized part and 1 GW shipping in the second half of 2026, plus custom EPYC CPUs.

AMD issued Meta a performance-based warrant for 160 million shares, about 10% of the company, at $0.01, vesting on shipment milestones and AMD's stock reaching $600. It is AMD's second mega deal of the same chips-for-equity shape, after OpenAI.

Anthropic published evidence that three Chinese labs ran coordinated distillation campaigns against Claude, generating over 16 million exchanges through roughly 24,000 fraudulent accounts — MiniMax over 13M, Moonshot over 3.4M, DeepSeek over 150k.

The post details chain-of-thought extraction, using Claude as an RL reward model, proxy and synchronized traffic, and coercion of "censorship-safe" answers, and says accounts were traced to specific researchers via request metadata. The accused labs have not confirmed the findings.