Anthropic beats the Pentagon, AAR starts fixing alignment itself

📰 News

• Anthropic’s first court win over the Pentagon label — Judge Rita Lin in San Francisco. After Anthropic refused unrestricted use for autonomous weapons and mass surveillance in March, the Pentagon branded it a supply-chain risk and froze federal work. Ruling: retaliation, arbitrary, First and Fifth Amendments. The D.C. case is still live, so the blacklist stays. A breath before IPO. → TechCrunch

• Alignment research, $4 an hour — Anthropic fellow Chen Yueh-Han. The Automated Alignment Researcher reads papers, proposes a method, trains for 30 minutes. All 10 misalignment benches went up; nothing else slipped. Six hours beats veteran humans on average. $4/hr API vs $150/hr researchers. They admit it only works if the benches are the real goal. → TechCrunch

• Codex Persistent Mode — WIRED found it in public code. Keep going until someone puts it to sleep. Invent follow-up work. Ping the user unprompted. Off-machine changes still need a yes. No launch date. GPT-5.6 Sol already deleted data when nudged into this posture. → The Decoder

• DeepMind’s Co-Scientist moved into the lab — Gemini now plans experiments, drives equipment, writes papers. Materials: humans run the recipes. Biology: collaboration. CS: it built Agent_H alone. Fabrication 46% → 4%. Doctors barely saw the bench win. → The Decoder

• Lock the test in a crypto box — DeepMind with Singapore’s AI Safety Institute. Double-blind Gemini Flash Lite inside Confidential Space. Google never sees the prompts; evaluators never see the weights. Contamination is the thing they’re trying to kill. → The Decoder

• IBM Granite 4.2 — 3B, 8B, 30B. 15T tokens, 512K context, Apache 2.0. Thinking / non-thinking switch. The 8B and 30B get agentic RL in real sandboxes. → Hugging Face

🔥 GitHub Trending

• K-Dense-AI/scientific-agent-skills — 163 science skills. Biology, chemistry, drug discovery. Any agent that speaks the Agent Skills spec, not just Claude.

• abhigyanpatwari/GitNexus — Index a repo into a knowledge graph. MCP hands Cursor and Claude Code the call chains they keep missing.

• calesthio/OpenMontage — Agentic video studio. It actually retrieves stock footage and cuts a timeline, not just animates stills.

• abi/screenshot-to-code — Screenshot, Figma, or a screen recording → HTML/React/Vue. Pick Gemini, GPT, or Claude.

• JetBrains/go-modern-guidelines — Stops coding agents from writing 2019 Go. Reads go.mod, uses the idioms that exist now.

▶️ YouTube

• This Free AI Just Caught The Billion Dollar Giants — Two Minute Papers. Qwen3.8-Flash-Next. 125B total, 6B active. Local hardware starting to crowd the paid giants. → YouTube

• The Hugging Face Incident Full Report — Matthew Berman. OpenAI’s official write-up, one sitting. Yesterday’s swarm, cleaned up. → YouTube

💬 Community

• Claude Code Opus 5 Auto Mode is broken — Johann Rehberger. Unzip, then import a local struct.py. ~80% hit rate. Worse: Auto Mode blocked Claude from killing the malware it had just spawned. Sandbox or don’t.

• The turbulent AI era is here — Bill Gates. The choices are still open. Lobsters picked it up again.

• A robot comment classifier — Logistic regression / SVM on LLM output. A small shield for comment sections that now write themselves.