Written by Admin Alex · Fact-Checked by M.Ali · Info Verified September 2026
We review and update this article regularly as new information becomes available.
TL;DR: Anthropic’s September 2026 threat intelligence report says a military-affiliated scientist asked Claude for help drafting a grant proposal to make the chikungunya virus more transmissible and better at dodging immune systems. The company also disrupted a Russia-linked group building an autonomous drone swarm that could pick its own targets, plus a wave of nation-state hacking campaigns. The report’s blunt conclusion: older AI models weren’t capable enough to meaningfully help with bioweapons. Newer ones might be.
A scientist with ties to a foreign military didn’t ask Claude how to build a bomb. He asked for help writing a grant application. The catch was what the grant was for: engineering the chikungunya virus, a mosquito-borne pathogen that already causes debilitating joint pain in outbreaks across Asia and the Americas, to spread more easily between people and slip past their immune defenses. Anthropic’s trust and safety team caught it, and it’s one of five biological weapons cases detailed in the company’s latest threat intelligence report, published this week.
The report reads less like corporate messaging and more like a field log of near misses. Beyond the chikungunya case, Anthropic says it stopped attempts involving avian influenza research aimed at mammalian adaptation, work with orthopoxviruses (the family that includes smallpox), and research into venom compounds and other biological toxins. The company’s own language is notably direct about the shift: “Older models were well below the threshold where they could meaningfully assist in bioweapons development. This is no longer a certainty with newer models.”
That’s not a line companies write lightly.
A drone that could pick its own targets
The report’s most unsettling section has nothing to do with biology. Freelancers linked to Russia reportedly used Claude Code, Anthropic’s coding assistant, to build an autonomous first-person-view drone system internally called DronDoc, also referred to as Serafim. According to Anthropic, the software gave a swarm of drones shared memory, terminal guidance logic, and the ability to select and engage targets, including a category the system explicitly labeled “person,” without a human confirming the kill decision. The report connects test-firing activity to Yemen.
Anthropic says it shut the project down once it was detected. But the fact that commercially available coding tools got this far, on what looks like a hobbyist-adjacent budget, is the kind of detail that should worry people well outside the AI industry.
Nation-states used Claude to hack, too
The cyber portion of the report reads like a greatest-hits list of 2026’s ugliest breaches. A Russian espionage campaign Anthropic tracks as GTG-20006 went after more than 20 Ukrainian and European targets, including government ministries, military intelligence units, and defense contractors. Separately, attackers linked to the ShinyHunters group scanned 1.8 million Android app packages hunting for exposed API keys, then pulled more than a terabyte of data, including millions of payment card records, from a single technology provider.
One especially brazen case: a group breached a software vendor and, within 34 hours, reached into systems belonging to 200 of that vendor’s downstream customers, netting tens of millions of airline passenger records along the way. Anthropic also flagged a Chinese university-linked operator that used Claude to help surface roughly a dozen possible zero-day vulnerabilities in major security products in a single month, and a separate campaign that targeted about 30 AI companies in just four days.
There’s a propaganda angle too. Anthropic says it disrupted nine influence operations running across six continents, including a France-based influence-for-hire network operating 70 fake news sites in roughly 20 languages, and a Malaysia-focused campaign built on nearly 1,000 fake social accounts.
Bottom Line
Anthropic gets credit for publishing this instead of burying it, and the level of technical detail here is genuinely useful for defenders. But read the report for what it actually says, not just how transparent it sounds: the company that builds Claude is telling you, in writing, that its own models are approaching the point where they can meaningfully help someone build a weapon. That’s not a hypothetical future risk anymore. It’s a current one, and every other frontier AI lab is racing toward the same capability threshold whether or not they’re publishing reports about it.
