Anthropic AI Misuse Report September 2026

In its September 2026 threat report, Anthropic details the disruption of numerous high-sophistication threat actors attempting to leverage Claude AI models to automate and scale offensive operations. Key identified actors include GTG-20006 (attributed to Midnight Blizzard/Russia), GTG-10007 (Chinese-speaking espionage group), and affiliates of the ShinyHunters criminal collective. These actors used AI to transition from traditional human-led operations to autonomous multi-agent frameworks capable of rapid vulnerability research, phishing, and automated malware refinement.

The report highlights a paradigm shift where AI has collapsed the labor and tooling gap between state-sponsored and individual operators. Techniques such as 'ClickFix' lures, DNS hijacking, and automated supply-chain theft were observed. Notably, threat actors are now targeting the AI supply chain itself, stealing API keys and session tokens to use as 'attack compute' for secondary intrusions. Commercial spyware vendors and state actors also misused AI for non-consensual surveillance, profiling of dissidents, and development of software for guided weapons and autonomous drone swarms.

Furthermore, the report documents industrial-scale 'illicit distillation' campaigns by major PRC-based AI labs, including Alibaba, DeepSeek, Moonshot, and Xiaomi. These entities harvested Claude's reasoning traces and chain-of-thought data to train their own models. This activity included 'replay attacks' that circumvented security controls and exposed sensitive user data, posing significant privacy risks and accelerating the proliferation of advanced AI capabilities to unauthorized actors.