Anthropic details how Claude was misused for surveillance and weapons
20 days ago
- Anthropic published a detailed report on misuse of its Claude models, covering cyberattacks, surveillance, influence campaigns, weapons development, and biological research.
- The report spans seven harm areas including cyber operations, influence operations, surveillance, conventional weapons, biological misuse, scams, and illicit distillation.
- In cyber operations, AI narrowed the gap between state hackers and lone operators, enabling single operators to conduct multi-victim campaigns in hours.
- Examples include a Russian espionage operation targeting Ukraine and Europe, and Chinese undergraduate students using agent swarms against 50 organizations.
- Surveillance misuse included a single consultant building a system monitoring 25 million SIM cards for Mali intelligence, and Chinese systems for scoring social media posts.
- Claude was used to write software for conventional weapons in China, Russia, and Yemen, including guided rockets, ballistic missiles, and drone swarms.
- Five cases of potential biological weapons support were documented, with difficulty in assessing intent; one involved gain-of-function research on chikungunya virus at a military institute.
- Illicit distillation by seven China-based labs (Moonshot, DeepSeek, Zhipu, etc.) extracted Claude's capabilities, with Moonshot relaying 300,000 requests via fraudulent accounts.
- Anthropic banned accounts, strengthened safeguards, and shared intelligence with authorities and industry partners.