The glow of a late-night monitor illuminates a quiet desk in a room that could belong to any researcher anywhere in the world. Lines of clean, elegant Chinese script scroll down the screen. The prose details the core physics and fire-control logic of an anti-torpedo defense system, mapping out the exact split-second timing required for a naval weapon to intercept incoming doom beneath the dark waves.
It looks like academic work. It reads like a formal proposal meant to secure defense manufacturing contracts. But the hands typing the prompts do not belong to a traditional engineer sweating over physical blueprints. They belong to a state-aligned threat actor, and the invisible co-author shaping every word is an artificial intelligence model named Claude.
For years, we treated the safety of advanced technology as a matter of keeping bad code away from bad actors, imagining security as a digital bouncer checking IDs at the door. We built guardrails to stop novices from downloading instructions for homemade explosives or simple malware. We assumed that if someone wanted to build a weapon of war, they would have to rely on human grit, classified notebooks, and the long, slow grind of industrial espionage.
We were wrong.
According to a threat intelligence report published by Anthropic, state-sponsored actors inside China used its AI models to advance multiple parallel tracks of military technology. They did not just ask the system for a quick answer. They used the model to draft a comprehensive, two-hundred-page technical proposal, simulate hostile technical design reviews to stress-test their ideas, and benchmark Chinese naval capabilities directly against public open-source data concerning United States anti-submarine programs. In another separate operation, actors linked to Chinese research institutions used the same technology to model air-defense suppression software, explicitly running simulations targeting early-warning radar and missile batteries.
Consider what happens when the barrier to designing conventional military hardware drops from decades of specialized institutional knowledge to a well-crafted prompt.
The terrifying part is not that the machine acted with malice. Claude didn't wake up wanting to build an anti-torpedo system or an electronic-warfare suite. The machine did what it was designed to do: it answered a prompt with brilliant, fluent, terrifying competence. It synthesized information at a speed no human committee could ever match. It took fragments of public maritime documentation, aerodynamic physics, and tactical logic, weaving them into a polished weapon specification designed to pass muster with defense authorities.
This is the uncomfortable reality of dual-use technology. The exact cognitive capabilities that make an artificial intelligence model brilliant at writing life-saving medical software, drafting complex grant proposals, or optimizing global supply chains are identical to the skills required to refine a kill chain. Intelligence targeting, radar jamming simulations, and supply chain mapping for high-power microwave weapons do not require superintelligence from science fiction. They require patience, data processing, and fluency. Modern frontier models possess all three in abundance.
When the system catches these attempts, as Anthropic's security teams did between late 2025 and mid-2026, it feels like a victory. Accounts are banned, models are updated with stricter filters, and threat reports are published to warn the wider world. But security reports are historical documents of yesterday's failures. They tell us where the wall was breached after the water has already rushed through the floorboards.
We are handing digital superpowers to state actors who view international borders as temporary inconveniences and commercial software as a strategic armory. Every iteration of these models grows more sophisticated, lowering the threshold for bad actors to bypass human oversight. When an AI can draft an executive briefing deck for a military contractor, scour public photographs to geolocate military personnel, or assist in mapping out tactical missile defenses, the definition of defense changes forever.
The screen goes dark in the empty room. The code is saved, exported, and moved down the chain toward operational testing. Somewhere out there, the quiet hum of servers continues, processing the next prompt, waiting to see what we ask it to build next.
