Summary
Fireship breaks down two overlapping AI safety stories from the same week: the viral resignation post from Jacob Coxin, a 27-year-old researcher who spent three years at OpenAI and Anthropic before quitting to accuse both companies of recklessly racing toward self-improving superintelligence, and Anthropic’s independently published 154-page threat report covering eight months of AI misuse the company says it detected and shut down. Coxin’s post reached 170 million views and 800,000 likes, and was publicly seconded by a current Anthropic employee who put odds of AI-caused human extinction above 10% within the next decade.
The threat report details seven categories of misuse with named threat actors. Russian hacking group Midnight Blizzard built a Claude Code workflow that automatically rewrites malware whenever antivirus software flags it. Two Chinese undergraduates ran an autonomous zero-day exploit foundry, loading firmware into a decompiler and running Claude-powered hypothesis-exploit-test loops around the clock. Shiny Hunters used Claude to mass-download and decompile 1.8 million Android APKs to extract hardcoded API keys, specifically targeting Claude and OpenAI credentials for resale.
The distillation section is the most expansive, accusing Chinese AI companies — Alibaba, DeepSeek, and Moonshot — of running hundreds of millions of Claude queries to harvest training data for competing models. Alibaba allegedly used fake accounts to make over one million requests per day to train Qwen. The report notes that distillation attacks were confined to Haiku, Sonnet, and Opus — none reached Fable or Mythos class models — which Fireship flags as evidence that frontier-model safeguards are improving.
📺 Source: Fireship · Published September 15, 2026
🏷️ Format: News Analysis
