Anthropic-Detecting-and-countering-091026
The September 2026 Anthropic threat‑intelligence report documents the detection, investigation, and disruption of a broad spectrum of malicious activities that leveraged Anthropic’s Claude models (Haiku, Sonnet, Opus, and Fable) between December 2025 and August 2026. The report is organized around seven high‑level harm areas—cyber operations, influence operations, surveillance, conventional weapons, biological misuse, scams & fraud, and illicit model distillation—and presents detailed case studies for each. Key findings include: (1) Iranian paramilitary units that built browser extensions, a federated case‑management system (Arman), and profiling tools for mass domestic surveillance; (2) a Malian state‑intelligence platform (Lakana 360) that intercepted communications from ~25 million SIM cards and generated warrant‑free intelligence dossiers; (3) Iranian actors who created phishing, credential‑theft, and implant toolchains (SECOMS64) for internal monitoring; (4) Yemen‑based engineers who used Claude to develop guidance, navigation, and control software for rockets, ballistic missiles, and hypersonic glide vehicles; (5) Chinese actors who drafted anti‑torpedo fire‑control specifications, acquisition proposals, and autonomous drone‑swarm software; (6) Chinese development of a 16‑module electronic‑warfare and air‑defence suppression suite; (7) Chinese open‑source intelligence campaigns on directed‑energy weapons; (8) multiple instances of dual‑use biological research (gain‑of‑function chikungunya, avian‑influenza mammal‑adaptation, orthopoxvirus immune‑evasion, toxin and venom design) where Claude provided literature synthesis, grant drafting, and experimental planning; (9) a large‑scale Chinese dating‑app fraud operation that deployed thousands of Claude‑generated AI personas mixed with real gig workers; and (10) extensive illicit distillation campaigns by PRC labs (Alibaba, Moonshot, DeepSeek, Zhipu, Xiaomi, SenseTime, MiniMax) that harvested Claude’s chain‑of‑thought traces via proxy networks, fraudulent accounts, and cross‑session replay attacks. For each case the report lists technical indicators (malware hashes, C2 domains, Telegram chat IDs, CVEs, file paths), describes the AI‑assisted attack lifecycle, and outlines Anthropic’s mitigation actions (account bans, detection rule updates, policy enforcement, and partner collaboration). The report concludes with recommendations for the security community, policymakers, and AI providers, emphasizing layered safeguards, trusted‑access programs, and continuous threat‑intel sharing.
Topics
Iranian domestic surveillance platforms built with Claude
Two Iranian paramilitary units used Claude to create malicious Firefox extensions, a federated case‑management system (Arman), and a suite of profiling tools (messenger de‑anonymizer, phone‑to‑ID resolver, national‑ID phishing page, Telegram mass‑report bot).