Anthropic Unveils Misuse of Claude for Surveillance and Weapons
Anthropic has revealed how its large language models, particularly Claude, have been exploited for malicious purposes, including surveillance, weapons development, and cyberattacks. In a detailed threat report, they detail cases involving actors from China, Russia, Iran, and Yemen.
Key Findings:
- Cyber Operations: Anthropic observes a decrease in the resources needed to conduct sophisticated cyberattacks, with lone operators now capable of targeting multiple victims simultaneously.
- State Surveillance: A Russian-speaking group, suspected to be Midnight Blizzard, used AI to evade security products during espionage campaigns against Ukrainian and European government targets.
- Weapons Software: Chinese undergraduate students built agent swarms targeting around fifty organizations, including a Southeast Asian government agency, retrieving citizen records.
- Biological Research: A single consultant based in Bamako, Mali, developed a surveillance system called Lakana 360, monitoring 25 million SIM cards across three mobile operators in Mali, bypassing legal requirements.
Anthropic emphasizes that they have detected and shutdown these misuse attempts, banning accounts and strengthening their safeguards. They also shared intelligence with relevant authorities and industry partners.
The Report’s Focus:
The threat report covers activity from December 2025 to August 2026, categorizing misuse into seven areas:
- Cyber operations
- Influence operations
- Surveillance
- Conventional weapons development
- Biological misuse
- Scams and fraud
- Illicit distillation
Anthropic notes that their Fable and Mythos-class models were not involved in any of these cases, except for one distillation case.