Overview
- Anthropic published a 154-page threat report on Sept. 11 that says it tracked misuse of its Claude models from December 2025 through August 2026 and disrupted many accounts tied to state-aligned and criminal actors.
- The company identified Iran-linked operations that automated profiling and targeting of hundreds in Israel and the Jewish diaspora and named several Iranian cultural and propaganda bodies as running state-aligned influence campaigns.
- Anthropic says Russia-linked developers used Claude to build software for an autonomous drone-swarm stack that could select human targets and issue detonation commands, and that China- and Russia-linked groups used the model for cyberespionage, malware and exploit development.
- The report describes five cases of 'biological misuse' where Claude assisted research with potential to aid harmful biological work, but Anthropic withheld country, institution and agent details while investigations continue.
- Anthropic says it banned accounts, removed relays, added detections and shared intelligence with governments and peers, and it urges session-level monitoring, mandatory incident reporting and stronger cross-industry governance to address evasion techniques such as prompt fragmentation, relays and large-scale distillation.