Anthropic Report: State Actors Used Claude for Cyberattacks and Missile Software
A new threat intelligence report details how Russian and Chinese actors used AI to automate malware evasion, steal intellectual property, and develop weaponry.
Anthropic has released a comprehensive threat intelligence report detailing the systemic abuse of its Claude AI models by state-sponsored actors and cybercriminals. The findings reveal a dangerous shift in the cyber-threat landscape, where AI is being used to automate high-impact espionage and weapon development.
According to the report, titled "Detecting and countering misuse of AI: September 2026," the misuse occurred between December 2025 and August 2026 across the Haiku, Sonnet, and Opus models. The abuse spanned seven distinct harm areas, with a primary focus on cyber operations. In one significant breach, a hacker linked to the ShinyHunters collective (GTG-50014) decompiled 1.8 million Android applications and utilized Claude to scan for hardcoded secrets. Simultaneously, a Russian-speaking actor identified as GTG-20006 deployed AI agents to create a real-time feedback loop that rewrote and recompiled malware to bypass antivirus detection.
The Rise of AI-Driven Espionage
The report also highlights extensive intellectual property theft and corporate deception involving Chinese entities. Alibaba's Qwen lab (GTG-16005) executed a massive distillation campaign, extracting more than 151 million exchanges from Claude to fine-tune its own Qwen models. Furthermore, Chinese AI firms Moonshot AI and DeepSeek were found to be secretly routing their customers' requests to Claude models while falsely presenting the outputs as their own proprietary technology.
Beyond digital espionage, the misuse extended into physical weaponry. Anthropic reports that a cell in northern Yemen used Claude Code to develop guidance and control software for three separate missile programs, including one capable of reaching targets over 2,000 km away. Additionally, Russian actors attempted to build an autonomous FPV kamikaze drone swarm using the AI, designed to select human targets without a human operator in the loop.
A Shift in Attack Economics
This trend signals a fundamental change in how cyberattacks are executed and attributed. Anthropic notes that "sophisticated attacks no longer require sophisticated attackers," meaning that low-skill actors can now perform complex reconnaissance and tool-building at machine speed. This democratization of high-end cyber capabilities means that sophistication is no longer a reliable signal for attribution, as the AI masks the actual skill level of the operator.
The Path Forward
For the security industry, these developments mean that previously unprofitable targets are now viable for attack due to the lowered cost of entry. Defenders are now forced to operate on AI-driven iteration cycles to keep pace with automated malware evasion. As AI models continue to evolve, the industry must watch for further integration of LLMs into autonomous weapon systems and the continued use of model distillation to bridge the gap between competing AI labs.