Anthropic Claude used for military targeting by Iran-linked actors
Anthropic disclosed that an Iran-linked account used Claude to compile U.S. military targeting data and naval system vulnerabilities by splitting requests into smaller steps to evade safety measures. The company stated reported outputs were not verified as used in actual attacks but represented a new form of intelligence collection using public AI systems.
What happened
Anthropic released a threat intelligence report disclosing that an Iran-linked account used Claude between December 2025 and August 2026 to compile scattered public information about U.S. military movements into a targeting manual and to catalog known vulnerabilities in naval ship systems.
The account used multi-step prompts designed to evade safeguards by breaking malicious requests into smaller steps that resembled legitimate research, achieving approximately 90% model refusal rates on direct harmful requests but unstable protection when requests were split across separate conversations with hidden final intent.
The account also used Claude to design Iran-based surveillance platform components combining license plate recognition and mobile device identifier interception, and other Iranian government entities used the model for propaganda material, false accounts, and social media analysis.
Reported impact
Anthropic stated the reported outputs were not verified to have been used in actual attacks. The report indicated that public military photos, transponder broadcasts, and commercial satellite imagery, when assembled by the AI, could form a more complete targeting package potentially endangering military deployments.
Index analysis
The incident demonstrates a new form of information exposure where adversaries use public LLMs as intelligence and programming assistants rather than conducting network intrusions. Anthropic found that model defenses are unstable when users distribute work across multiple conversations and obscure ultimate purpose.
Governance controls
Anthropic blocked the account, implemented automated detection based on observed behavior patterns, and shared threat intelligence with government and private sector partners. The company stated Iran was outside its service region and the user bypassed geographic restrictions using proxy servers, account rotation, and third-party forwarding.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.