AI giant Anthropic has released a chilling safety report detailing how state-linked actors and criminals attempted to use Claude to engineer bioweapons, autonomous drone swarms, and ballistic missiles.
- State-backed groups attempted to use Claude for gain-of-function research on viruses like Chikungunya and Orthopoxvirus.
- AI was leveraged to develop software for hypersonic glide vehicles and autonomous 'kamikaze' drone swarms in Russia and Yemen.
- Anthropic disrupted multiple espionage and mass surveillance attempts between December 2025 and August 2026.
In a comprehensive and alarming safety disclosure, Anthropic has revealed the dark side of Large Language Model (LLM) utility. The company documented a series of sophisticated attempts by state-sponsored actors, spyware vendors, and financially motivated criminals to weaponize the Claude AI ecosystem. These activities, spanning from late 2025 through mid-2026, highlight a growing global arms race where AI is no longer just a productivity tool but a target for strategic military exploitation.
The Biological Threat: Engineering Pathogens
The most concerning findings involve biological research. Anthropic identified instances where users attempted to use the AI to draft grant applications for gain-of-function research. Specifically, one user sought to engineer mutations in the Chikungunya virus to make it more harmful—a request Anthropic believes was destined for a military research institute. In another instance, the Opus 5 model was nearly used to facilitate research on Orthopoxvirus, the family of viruses responsible for smallpox.
The complexity of these threats lies in the "dual-use" nature of biological science. Legitimate medical research often mirrors the steps required to create a weapon. To combat this, Anthropic implemented biological safety classifiers and proactive account sweeps to identify users circumventing regional restrictions.
Why This Matters
BozokMedia analysis shows that the transition from 'chatbots' to 'coding agents' (like Claude Code) has drastically lowered the barrier to entry for high-tech weaponry. When AI can replace human guidance, navigation, and control engineers, the speed of weapon development accelerates exponentially, bypassing traditional academic or institutional oversight.
"The risk is no longer about a comic-book villain asking for a doomsday device, but about nuanced, academic-style requests that mask lethal intent."
Conventional Weaponry and Autonomous Warfare
Beyond biology, the report details terrifying applications in conventional warfare. In Yemen, a cell of actors used Claude Code to develop software for a guided rocket and a hypersonic glide vehicle with a range exceeding 2,000 kilometers. Meanwhile, in Russia, an operator linked to the 'DronDoc' project used the AI to engineer a first-person-view (FPV) kamikaze drone swarm capable of selecting human targets without a human in the loop, using scraped combat footage for vision training.
| Region | AI Application | Target/Weapon |
|---|---|---|
| Yemen | Guidance Software | Hypersonic Glide Vehicle |
| Russia | Autonomous Logic | Kamikaze Drone Swarm |
| China | Military-Industrial | Conventional Weaponry |
Anthropic noted that while some efforts failed—such as the Yemeni rocket test—the ability of these actors to route around safety controls using third-party evasion platforms remains a critical vulnerability for all AI providers.
Frequently Asked Questions
Q1: Did any of these attempts succeed in creating a weapon?
While some software was developed and tested (like the Yemeni rocket), Anthropic claims most high-risk biological and software requests were blocked or limited to weaker model tiers.
The Haiku, Sonnet, and Opus models were frequently used, while the more advanced Fable and Mythos-class models were largely avoided, except in one specific case.
