SAN FRANCISCO - The real threat of artificial intelligence doesn't look like a sci-fi movie about rogue robots. It is much quieter: a scientist at a foreign military lab asking a chatbot how to make a virus deadlier, or a militant group using AI to troubleshoot the steering on a ballistic missile.
And for months, these haven't been theoretical risks. They were real operations.
On Thursday, September 10, 2026, AI developer Anthropic released a 154-page threat intelligence report detailing how state-sponsored spies, criminal syndicates, and militant factions targeted its Claude models between December 2025 and August 2026. The report offers the clearest picture yet of how malicious actors are evolving. They are no longer just asking chatbots to write basic phishing emails; they are embedding AI into the heart of their military and cyber operations, treating it less like a search engine and more like an automated engineering workforce.
According to Anthropic, none of these operations managed to breach its newest, most heavily guarded systems, such as the Mythos and Fable-class models. Instead, attackers relied on older versions like Claude Opus 4 and Sonnet 4.5, hiding their tracks behind third-party resellers and digital proxies.
Yet, even with older technology, the results were alarming.
The Biological Gray Area
Some of the most concerning incidents involved professional scientists conducting biological research. The terrifying reality of biology is that the exact same data used to create a life-saving vaccine can also engineer a devastating bioweapon.
Because of this "dual-use" nature, Anthropic stressed that it couldn't definitively prove the scientists meant to cause harm. But the nature of their requests (and the identities of the people making them) triggered immediate alarms.
“You are not seeing someone in a comic book kind of way say, 'Hey, I want to build a biological weapon to kill everybody,'” Jacob Klein, Anthropic’s head of threat intelligence, told The New York Times. “It’s an incredibly nuanced situation. What we don't know is if the research was meant to be weaponized.”
In one case from May, a researcher linked to a foreign military institute used Claude to help draft a grant proposal on the chikungunya virus (a mosquito-borne illness that causes agonizing joint pain). The researcher was actively seeking genetic mutations that would make the virus spread faster and kill more easily in mammals.
In another incident, users took just an hour to generate a complete research proposal on how orthopoxviruses (the family that includes smallpox and mpox) could be engineered to evade the human immune system. Anthropic also caught researchers using Claude to catalog venoms capable of causing sudden respiratory failure, and another scientist who spent weeks figuring out how to adapt the highly lethal bird flu so it could easily infect mammals.
A Cat-and-Mouse Game in Cyberspace
The report also outlines how international hackers are upgrading their arsenals.
Anthropic tracked one major espionage group tied to Midnight Blizzard, a notorious Russian intelligence hacking unit. Instead of just using AI to write malicious code, the group built a fully automated feedback loop. If security software caught their malware, the hackers simply fed the error alerts back into Claude. The AI analyzed the failure, rewrote the code to hide it better, and deployed it again—repeating the process until the malware slipped past defenders.
The same Russian-linked group compromised the guest Wi-Fi providers at international hotel chains. When diplomats or defense executives logged on from their rooms, they received fake, AI-generated software updates designed to steal their passwords.
In a separate cyberattack, hackers deployed autonomous "AI agents" (programs that operate without human supervision) to breach a corporate cloud provider. In just 34 hours, the AI navigated the network, stole thousands of security tokens, and compromised roughly 200 different customer organizations.
AI on the Battlefield
Perhaps the most startling revelation is how AI is helping build physical weapons. Anthropic intervened in six operations involving hardware, ranging from drone swarms to missile systems.
The most extensive operation took place in northern Yemen, an area controlled by Houthi militants. An engineering cell there used Claude as a substitute for human developers across three major weapons programs: a guided rocket powered by standard smartphone chips, a ballistic missile with a 2,000km range, and a highly advanced hypersonic glide vehicle.
The engineers essentially built a virtual assembly line, assigning different AI "agents" specific tasks. One modeled aerodynamics, another wrote navigation software, and a third checked the code for errors. When a physical rocket crashed during a test launch, the team fed the crash data back into Claude within hours to troubleshoot the failure.
While Anthropic noted there is no evidence the group successfully deployed a working missile before their accounts were banned, the attempt marks a dangerous new frontier.
In a similar case, Anthropic caught a freelance team in Russia using Claude to program small kamikaze drones. By training the drones on combat footage, the team taught them to identify human shapes and detonate autonomously, removing the need for a human to pull the trigger.
The Regulatory Reckoning
The Anthropic report exposes a glaring gap in global security: international laws are designed to regulate the sale of physical weapons and chemicals, not lines of code and digital intelligence.
Today, threat actors easily bypass geographic bans using virtual private networks (VPNs) and shell companies to access Western AI systems. Now, governments in Washington and Europe are scrambling to adapt.
According to trade officials, regulators are seriously considering forcing AI companies to adopt the strict "Know Your Customer" laws currently used by banks. If passed, companies like Anthropic and OpenAI would be legally required to verify the exact identity of anyone trying to use their most powerful models, treating access to cutting-edge AI with the same caution as the sale of military hardware.