Zvi on Anthropic's misuse report: Chinese labs' distillation is the real threat
Don't Worry About the Vase (Zvi) · rss · 2026-09-15
Zvi Mowshowitz breaks down Anthropic's threat intelligence report covering December 2025–August 2026 across seven harm areas. Key takeaways:
- Most malicious actors remain unsophisticated; AI's main effect is turning 'bad at job' into 'good at job.'
- Fraudulent distillation is the report's most consequential threat: all top Chinese labs allegedly tried to distill Claude, with DeepSeek, Moonshot and Xiaomi sending real user queries — Moonshot reportedly passed outputs to users as Kimi results; Zhipu failed against Fable and moved to Opus.
- Notable cases: Midnight Blizzard-linked Russian espionage automating evasion and phishing against Ukrainian targets; ShinyHunters 'vibe hacking'; a Chinese-speaking op running Claude agent swarms for vulnerability research against 50 organizations; a Russian-speaking actor prompt-injecting 30 companies hunting pre-release Claude access; nine influence-op networks built with Claude, though defenders are mostly winning.
- Zvi also flags Trump declaring AI existential risk a 'hoax' as very bad news.
More from Companies & People
- Cohere CEO alleges frontier labs are rigging AI regulation into a regulatory moat — sourdub · 2026-09-16
- Merge's 'My Claude's Promiscuous' billboard goes viral with 275K views — shensi · 2026-09-16
- After welcoming Dario at Dreamforce, Benioff's feed draws jab: if you truly believe in 10% extinction risk, why sell AI into B2B SaaS — SumitGup · 2026-09-16
- Latham & Watkins, No.2 US Law Firm, Buys Nvidia Hardware to Fine-tune Open Weights In-house — MikeBirdTech · 2026-09-16
- Andrew McAfee on 'Geek Doctrine': how Silicon Valley ran circles around incumbents — amcafee · 2026-09-16
- Jensen Huang: AI Safety Is an Engineering Problem, No New Laws Needed — DavidSacks · 2026-09-16