Countering misuse of AI: September 2026 / Anthropic

Anthropic's Threat Intelligence team has published a report on malicious uses of its Claude models that it disrupted between December 2025 and August 2026. It builds on earlier threat reports from March, August and November 2025. The report presents selected case studies in seven areas of harm, among them cyber operations, influence campaigns, surveillance, fraud, biological misuse, weapons development and distillation. Anthropic says it disrupted each case, tightened its safeguards, and passed what it learned to authorities and industry partners when suitable.

The actors include suspected state-sponsored groups, profit-driven criminals, commercial spyware vendors, state propaganda bodies and politically motivated individuals. Haiku, Sonnet and Opus models were involved. Apart from one illicit distillation case, no misuse involved Fable or Mythos-class models. Anthropic says the cases are the most notable and novel it has found, published out of a duty to disclose.

The cyber section argues that sophisticated attacks no longer need sophisticated attackers, because AI has narrowed the labor and tooling gap between well-resourced state operations and individuals. It also says AI's role has become more autonomous, with multi-agent frameworks carrying out reconnaissance, exploitation and data theft while humans choose targets and review stolen material. One case, tracked as GTG-20006 and described as Russian espionage, used AI workflows to rebuild malware whenever security products detected it, targeting more than 20 organizations, mostly in Ukraine and Europe. Another, GTG-50014, involved suspected affiliates of the ShinyHunters collective who used AI to speed up scanning, credential theft, data theft and extortion, including a supply-chain breach reaching roughly 200 downstream customers.

Anthropic says it will keep evolving its safeguards and hopes the findings help other developers, governments and civil society recognise similar threats.