AI firm reports misuse of Claude in weapons development and cyberattacks
Anthropic says it blocked attempts to use its Claude AI for missile guidance software in Yemen and for state-linked cyber-espionage. The company reported that a Russian-linked group used automated AI workflows for phishing and data theft, and a Chinese operation was also disrupted. Anthropic banned the accounts and shared threat information with partners.
Anthropic's report describes how operators in Yemen bypassed safety filters by splitting missile-guidance coding tasks across separate sessions and assigning distinct roles to different model instances. Although a test-fire reportedly failed, the firm confirmed that some requests evaded internal checks before the accounts were terminated.
The report also details state-linked operations, including a Russian group automating phishing and data theft against European and Ukrainian targets, and a Chinese student group orchestrating network intrusions. Additionally, Iranian and Chinese-aligned actors leveraged Claude for influence campaigns and targeted recruitment, with one operation using the model to translate and draft messages in regional dialects for infiltration efforts.
This disclosure could heighten scrutiny of AI safeguards, as adversaries increasingly adapt to evade them. Governments and defense contractors may face elevated risks from automated espionage and weapons development, while the public could see accelerated calls for stricter international AI governance. The use of models for targeted recruitment and influence operations may also complicate efforts to protect vulnerable communities, potentially reshaping how tech firms balance openness with security in conflict zones.