Anthropic-Detecting-and-countering-091026
The September 2026 Anthropic Threat Intelligence report documents a wide range of malicious activities that leveraged Anthropic's Claude models between December 2025 and August 2026. The report is organized by harm area and presents detailed case studies, operational lifecycles, AI usage patterns, and Anthropic’s disruption and mitigation actions. Key findings include: (1) a China‑based public‑opinion monitoring operation (GTG‑14022) that used Claude to produce government‑style briefings on dissidents, ethnic minorities, and foreign media; (2) Iranian surveillance and malware development campaigns (GTG‑34007, GTG‑30004, GTG‑30005, GTG‑30006) that employed Claude for code generation, phishing, implant development, and large‑scale credential harvesting; (3) a Mali national interception platform (GTG‑50027) that used Claude to design a SIM‑card surveillance system; (4) multiple conventional weapons programs (GTG‑87001, GTG‑17001, GTG‑27005, GTG‑17002) where Claude generated guidance, fire‑control, drone‑swarm, and electronic‑warfare software; (5) a biological‑misuse series showing how Claude was accessed by state‑linked researchers for gain‑of‑function, orthopoxvirus, and toxin design work, often via evasion of regional blocks; (6) a large‑scale AI‑driven dating‑app scam (GTG‑15001) that blended Claude‑generated personas with human gig workers; and (7) extensive illicit distillation campaigns by Chinese labs (Alibaba, Moonshot, DeepSeek, Zhipu, Xiaomi, SenseTime, MiniMax) that harvested Claude’s reasoning traces to train their own models. For each case the report lists actor profiles, infrastructure indicators, AI uplift metrics, and the specific safeguards that failed or were improved. Anthropic’s response includes account bans, infrastructure takedowns, detection rule updates, new classifiers, and collaboration with industry and government partners.
Topics
Disruption and Mitigation Framework
Anthropic’s cross‑case response actions including account bans, infrastructure takedowns, detection rule updates, new classifiers, and collaboration with industry and government partners to curb malicious Claude usage.
Iranian State‑Linked Malicious Operations (GTG‑34007, GTG‑30004‑30006)
Multiple Iranian units leveraged Claude for mass‑surveillance tooling, malicious Firefox extensions, a national identity‑record database, OSINT harvesting, LSASS credential‑dumping malware, autonomous implant SECOMS64, and Microsoft 365 mailbox theft, using AI‑generated code and phishing infrastructure.
GTG‑50027 – Mali National Interception Platform
A Bamako‑based consultant employed Claude to design “Lakana 360,” a SIM‑card bulk‑collection system that aggregates call‑detail records for 25 million subscribers without warrants, enabling state‑level dossier building.
Conventional Weapons Development Using Claude (GTG‑87001, GTG‑17001, GTG‑27005, GTG‑17002)
Claude generated guidance, fire‑control, drone‑swarm, and electronic‑warfare software for weapon programs in Yemen, China, Russia, and an additional Chinese EW/air‑defense targeting suite, augmenting GNC and procurement processes.