It turns out that artificial intelligence, the same tech promising us flying cars and instant translations, might also be a tool for folks dabbling in the scariest kind of science. Anthropic, the big brain behind the Claude AI chatbot, has dropped a report saying it sniffed out five cases where users tried to get its AI to help with research that could, you know, lead to biological weapons.
These dodgy dealings weren't ancient history; they were spotted between December 2025 and August 2026. The company detailed these in its latest threat intelligence report, released this past Thursday. We're talking about research areas that sound straight out of a thriller novel: the chikungunya virus, the nasty highly pathogenic avian influenza (that’s bird flu to you and me), orthopoxviruses (which includes things like smallpox), and even venom peptides and toxins.
Anthropic says they’ve already shut down the accounts involved and have beefed up their own AI's defences based on what they learned. But here's the kicker: they're not accusing anyone of actually planning to make bioweapons. In fact, they're keeping quiet about the names of the researchers, their labs, and even the specific nasty agents they were looking into. They say outing them could actually put them in danger. Honestly, distinguishing between legitimate science and a potential doomsday plan is getting harder by the day.
“Biological misuse is one of the most serious risks of frontier AI models,” Anthropic warned. It’s not hard to see why. AI could potentially be used to make existing superbugs even more dangerous, or maybe even cook up brand new ones. The company pointed out the crazy paradox: the same information that could build a weapon could also help create a vaccine or a cure. It's a classic case of dual-use technology, a double-edged sword.
Let’s look at how this played out. In one instance, Anthropic’s AI safety filters caught a request to help draft a grant application for research that involved tinkering with the chikungunya virus – what they call “gain-of-function” research. Further digging revealed that some clever folks were using a reseller platform to get access to Claude in places where Anthropic doesn't officially operate. If Claude said no, this platform apparently had a backup plan: sending the request to other AI models with fewer rules.
Anthropic didn't just twiddle their thumbs. They banned the accounts and worked with partners to shut down these sneaky relay networks. They also shared their findings with other AI companies and government folks. Then there was the case of a researcher outside the US who spent weeks using Claude to plan research on highly pathogenic avian influenza. They were specifically looking at how it could adapt to infect mammals.
This researcher used Claude for everything from planning the study and analysing data to figuring out which experiments to prioritise and even for some editing.
Anthropic did manage to limit this researcher to weaker versions of Claude, so the AI’s help was mostly on the administrative side, sparking ideas, and designing the study. Still, it proved that serious research programs focused on pathogens with serious pandemic potential were definitely active. Another incident involved a grant application for orthopoxvirus research at a lab linked to the state. A reseller service, serving over a dozen customers, used Claude to help draft this application. It focused on how these viruses mess with the immune system.
Get this: the whole thing was drafted using Claude Opus 5 in about an hour.
And then there's the research into venoms and toxins. One researcher apparently used Claude to build a whole map of venom toxin peptides. They even created a system to fine-tune toxin characteristics, all with the stated aim of developing new medicines. In another related case, a researcher was using Claude for projects involving the computational redesign of toxins. This was part of a national research program.
This person even went so far as to deliberately hide the identities of some toxins and viral proteins in their progress reports. Both these accounts got the boot in May 2026 for breaking Anthropic’s policy on supported regions.
These cases really hammer home how tough it is to keep tabs on AI-assisted biological research. So much of it is “dual use” – it can be used for good, like fighting disease, or for bad, like creating weapons. Anthropic, though, is quick to say this isn't proof that Claude is actively helping to create biological threats right now. Instead, they see it as evidence that significant research that could be used for harm is happening. It's associated with states that raise concerns.
This report is part of a bigger picture, with Anthropic being more open about how its AI systems are being misused. They are navigating a tricky path, trying to enable beneficial scientific discovery while preventing catastrophic misuse. The challenge isn't just for Anthropic; it's for the entire world as AI gets more powerful and the lines between progress and peril blur.