Claude did not invent scorpion antivenom

Citrini Research, a firm with no biology background, set out to use an unreleased Claude model to make a new scorpion antivenom. Anthropic lifted its biosecurity guardrails, hosted the team, and covered close to $50,000 in compute. Claude generated nearly 320,000 candidate proteins in 64 hours, the team narrowed that to 211 designs, and outside labs found 17 that bound to four of five targeted toxins. None bound more than one.

That is not a drug, and it is not close. Experts who spoke to The Verge said binding is the relatively easy step; neutralizing a toxin safely inside a living body is the hard one, and nobody has tested that yet. With Anthropic weeks from its IPO and pushing into life sciences, how this kind of result gets framed matters for investors, patients and scientists alike.

Let's give the experiment its due first. A small, non-expert team pointed a model and hundreds of subagents at a real biomedical problem and got testable binders in under three days. The Danish researchers The Verge consulted called the autonomous discovery process the most novel part, and we agree. That speed, if it holds up, is the kind of useful AI we want to see in labs.

But the framing outran the evidence, and that is the problem we keep coming back to. A blog post titled around creating a new drug, a declared success of sorts, a pitch to paying subscribers about which companies stand to benefit: that is a market narrative built on a step one experts describe as close to solved. One professor told The Verge that existing screening methods already find plenty of hits at manageable cost without AI. The gap, in his words, is the translation into medicine. When an investment shop with a history of market-moving fiction says the AI part was free, while Anthropic quietly paid the bill, we should read the whole thing as promotion first and science second.

The guardrails piece bothers us more than the hype. Citrini hit Claude's biosecurity restrictions, called around to lab contacts, and Anthropic agreed to drop them for a more capable unreleased model, under supervision in its offices. The team avoided candidates resembling known toxins, and supervised access is a defensible model. Still, the path from blocked user to unlocked frontier bio model ran through personal contacts and a company with a commercial reason to want a splashy result. If that becomes the normal route, guardrail exceptions will track who has connections and a good story, not who has the expertise and oversight.

Where this leads if it scales is easy to picture. More AI drug-discovery announcements that stop at binding, more retail money chasing the testing and supply companies those announcements highlight, and more pressure on the real translational science that takes years. The tools are getting good. The claims need to wait for the wet lab.

Anthropic's own account was more measured, describing the candidate generation and filtering. Citrini paid roughly another $150,000 for binding tests at Adaptyv Bio and Twist Biosciences. The Verge noted it was unclear why Anthropic's own California wet lab was not used. Citrini and Anthropic did not respond to The Verge's requests for comment.