They published the kill chain

Anthropic dropped its densest threat-intel report yet. Lab researchers kept the slowdown chorus going. Musk called the Coxon wave a setup — treat that as his allegation.

Anthropic today (~6:13pm London): published Detecting and countering misuse of AI: September 2026, billed as their most detailed threat-intel report yet. Covers Dec 2025–Aug 2026 cases they say they disrupted across cyber, influence ops, surveillance, scams/fraud, biological misuse, conventional weapons, and distillation (mostly Haiku/Sonnet/Opus; not Fable/Mythos except one illicit distillation case). Standouts: suspected Russian Midnight Blizzard–linked GTG-20006 using AI to auto-rebuild malware when detections hit, plus hotel WiFi DNS hijacks and drone-supply targeting; ShinyHunters-linked smash-and-grabs (incl. vibes-hacking and stolen customer API keys as attack compute); Chinese-speaking GTG-10007 running agentic exploit foundries; and a slate of influence ops (Russia/Iran/commercial influence-as-a-service). They say every named op was disrupted and intel shared with partners. anthropic.com/threat-intelligence-report-september-2026

Labs’ own researchers keep calling for a slowdown (today). CNBC: OpenAI safety staffer Julie Steele (“we need to slow down”), Anthropic’s Samuel Marks and Anna Wang, OpenAI’s Jasmine Wang on RSI danger — all amplifying the Coxon/Hubinger wave. Geoffrey Hinton told BBC Newsnight a >10% decade extinction risk “seems a not unreasonable estimate.”

Musk vs Coxon (overnight into today). Musk: “Seems like a setup,” then “psy op,” and that ~123M views in a day from a near-new account “has never happened.” Coxon pushed back that he’s real. WaPo today: Coxon is now a right-wing political target. Treat the “setup” claim as Musk’s allegation, not established fact.

Light / already in the morning brief: Christiano’s “most people could die” / OpenAI Foundation board seat still circulating (Guardian/BI); Anthropic’s fourth Claude cyber-eval incident and OpenAI’s Lehane federal-safety push were morning items — skip rehash. No fresh Altman / Hassabis / Karpathy / Sutskever / Huang product news in the window (NVIDIA did flag Zagreb driverless robotaxi tests with PonyAI/Verne).