Full Breakdown
Anthropic Discloses Widespread Misuse of Claude AI by State-Linked Actors
By Drooid · · How we work
Core Event: Weaponization of Claude Across Multiple Threat Actors
Anthropic released a 154-page threat-intelligence report on September 10 2026 describing how its Claude model was employed between December 2025 and August 2026 for cyber-espionage, weapons development, influence campaigns, surveillance, biological-research assistance, and illicit model-distillation. The company says it blocked dozens of accounts, banned the users, and shared intelligence with government and industry partners.
Background & Context
Claude is a publicly accessible generative-AI assistant offered by Anthropic. The firm’s Threat Intelligence team monitors API usage for “uplift” – rapid gains in speed, scale, or depth that indicate malicious automation.
Key Actors & Operations
- Midnight Blizzard – a Russian-state-linked hacking group used Claude to automate reconnaissance, phishing, malware rebuilding, and data exfiltration against more than 20 Ukrainian and European targets, including ministries and drone-technology firms.
- Iran-linked actor – employed Claude to compile targeting handbooks for U.S. Navy vessels and to surveil 6,388 Iranian dissidents, producing propaganda for “explanatory jihad.”
- UAE-directed influence operation – actor “Deadshot” used Claude to generate reports, ghost-write UN-Human-Rights-Council testimonies, and profile 18 European-Parliament members while omitting any mention of the UAE.
- Chinese AI labs – Alibaba, Moonshot, and DeepSeek illicitly distilled Claude outputs, generating over 151 million exchanges (Alibaba) and millions more from the other labs between May and July 2026.
- Yemen cell (Houthis) – leveraged Claude Code to develop guidance software for a multi-stage ballistic missile with a claimed range above 2,000 km; a test-flight failed, prompting further AI-assisted troubleshooting.
- Mali surveillance platform – a consultant built “Lakana 360,” monitoring roughly 25 million SIM cards across the country’s three mobile operators using Claude as the primary engineering tool.
Data & Statistics
- 20+ Ukrainian and European organizations were scanned or compromised.
- 300,000+ national-identity records and 500,000+ commercial-registry entries were stolen from a North-African agency.
- 4,700 Claude-powered AI personas exchanged 2.36 million messages with 25,000 users in a two-week April 2026 dating-app scam.
- 151 million exchanges (Alibaba) and 23 million (Moonshot) were recorded during distillation campaigns.
Official Statements & Responses
Anthropic says it banned every account linked to the misuse, enhanced detection classifiers, and shared findings with law-enforcement and industry partners. New safeguards restrict dual-use biological queries on newer Claude models and add identity-verification steps for accounts showing abuse signals. The company notes that “disruption” primarily means account bans and does not guarantee recovery of stolen data.
Criticism & Opposition
Some analysts argue the safeguards remain insufficient to stop AI-enabled threats.
Verbatim Quotes
- “One of the [Midnight Blizzard] operators is a Russian speaker using the handle ‘JackPoterz’ whose tradecraft and targeting are consistent with Russian state-nexus espionage,” — Anthropic senior researcher
Conflicting Reports & Gaps
Attribution of the Russian campaign varies: some sources label the group “Midnight Blizzard” and tie it to the Russian Foreign Intelligence Service, while others describe it more generally as a “Russia-linked hacking operation.”
Timeline
- December 2025 – Start of the observation window.
- September 10 2026 – Anthropic publishes the threat-intelligence report.
- August 2026 – End of the observation window.