Anthropic blocks state-linked AI misuse for weapons, cyberattacks, and surveillance
Anthropic has disrupted attempts by state-linked groups in Russia, China, Iran, and Yemen to use its Claude AI models for cyberattacks, weapons development, and surveillance, per a new threat report. The company also flagged five cases where researchers used Claude in ways that could support biological weapons development, though it did not assert malicious intent.
Bottom line — Anthropic says AI is automating large parts of cyber operations, reducing the need for skilled human operators.
Go deeper (7)
- A Russia-linked group consistent with Midnight Blizzard used Claude across phishing, hotel Wi-Fi hijacking, and WhatsApp takeovers targeting Ukrainian government and military organisations, per Anthropic. The attackers built an automated system that rewrote malware until security tools stopped detecting it.
- In northern Yemen, operators attempted to use Claude to write guidance and flight-control software for a guided rocket and long-range ballistic missile, according to Anthropic. The company said it had no evidence a working weapon was produced.
- Anthropic linked three accounts to Chinese municipal security services that used Claude to profile overseas activists and organisations as part of surveillance and transnational-repression efforts, per the report.
- The company also accused seven China-based AI labs of illicit distillation, with Alibaba-linked accounts generating over 151 million Claude interactions between May and July using more than 3,500 fraudulent accounts, per Anthropic.
- Anthropic said it blocked a grant application for gain-of-function research on the chikungunya virus aimed at enhancing transmissibility and immune evasion, per the report. The company noted such research could also be used to develop vaccines.
- The report comes days after former Anthropic researcher Jacob Coxon resigned, accusing the company and OpenAI of 'racing straight to self-improving superintelligence and gambling with our lives,' per his posts on X.
- Anthropic said its older models were below the threshold for meaningfully assisting dangerous biological research, but newer models like Claude Fable 5 have stronger safeguards restricting dual-use biological queries, per the report.