Why we still ask a horse to save us from scorpion venom

Imagine a tiny key.

A scorpion carries it. One sting, and the key enters the body.

Inside, that key opens doors that were meant to stay shut. The body goes haywire.

To make the antidote, we go through a horse. We inject it with a little venom, we wait months, we draw its blood.

It works. It costs a lot. And where the scorpions live, there is often none left.

This autumn, a different question was put to an artificial intelligence: can you invent a cap for this key?

A cap that sits on top and stops it turning.

The machine drew 318,484 of them. Then it chose the ones worth making for real.

Of the 211 made and tested in a laboratory, 17 hold on to the key. Like a mark at school: 17 successes out of 211 attempts, and two genuinely excellent papers.

This is exactly what has just happened. The artificial intelligence is called Claude, the keys are five scorpion toxins, and the caps are proteins that did not exist before.

You will see none of it in a pharmacy. Not this year, not the next. Nothing has yet been tried in an animal, let alone in a person.

But something has changed: every year around 1.2 million people are stung by a scorpion and more than 3,250 die, mostly children. Nobody gets rich treating them. That is precisely why this work was moving slowly.

So the question is not whether the machine is clever. It is who decides which doors it tries to open.

For those who want the detail, here is what happened.

What happened, in plain words

Scorpion venom contains neurotoxins: very small proteins that attach to the channels our nerves use to pass their signals, and throw them out of order. That disorder is what kills.

Conventional antivenom has been made the same way for more than a century. Venom is collected from thousands of scorpions, injected into horses, their bodies are left to produce antibodies, and those antibodies are then purified from their blood. The process takes months, costs a great deal, and each antivenom protects against only a small number of toxins.

In October 2026, Anthropic's Science team published the results of a campaign run with the research firm Citrini Research. The principle: ask Claude to design from scratch proteins able to attach to five scorpion neurotoxins from Mexico, Brazil, North Africa and the Middle East. Designing "from scratch" means inventing the shape of the protein, not copying it from an existing antibody.

The sequence, as the organisation describes it: one opening instruction, then four human follow-ups, across a 64-hour campaign. After the first instruction, Claude worked alone for 24 hours. It set its own quality thresholds, chose the toxin structures and the areas to aim at, and selected 180 candidates to be made. The human team added 31 more, directed the work and ordered the synthesis. That is 211 proteins actually produced and tested by two independent laboratories, Adaptyv Bio and Twist Bioscience.

The result: 17 attach to their target, seven of them to the Cn2 toxin, the one from the Mexican scorpion. The best two hold on very tightly. Two scientists outside the project, Susana Vázquez Torres (CNIO, Madrid) and Timothy P. Jenkins (Technical University of Denmark), reviewed the work and call it a promising first step.

What did not work counts just as much: the proteins meant to neutralise several toxins at once all failed. And the tests were run on toxins reproduced in the laboratory, not on real venom.

Artificial intelligence and ethics: why this case is not like the others

What matters is not the technical feat. It is the choice of target.

Scorpion envenoming is one of what the World Health Organisation classes as neglected tropical diseases. Neglected means: they kill, but the people they kill cannot afford treatment. No laboratory builds its profitability on that.

When the design stage of a medicine goes from several months to a few dozen hours, the arithmetic changes. What was not profitable becomes feasible. That, and not raw performance, is where artificial intelligence and ethics genuinely meet.

The warning

Three questions, without alarmism.

Who will own the result? Designing the molecule is the part that has just become fast. Producing it, stabilising it, carrying it refrigerated to a rural clinic and selling it at a bearable price is the part that has not moved an inch. A brilliant protein locked behind an unreachable patent saves nobody.

Can the same skill be used to cause harm? Knowing how to design a protein that attaches precisely to a biological target is a double-edged capability. The subject is not addressed in the publication. It should be, and it deserves an outside, independent look.

Will we mistake a step for a victory? Between 17 proteins that attach in a tube and a vial on a clinic shelf lie cell tests, animal tests, toxicity, stability, clinical trials and funding. The organisation itself writes that it does not claim to have produced an antivenom. It should be believed on that point as much as on its figures.

The hope

It is real, and it is twofold.

First, the work was reviewed by outside scientists, confirmed by two independent laboratories, and the data was made public so that other teams can use it. That is the opposite of a closed announcement.

Second, the precedent matters more than the result. If a handful of hours of computation is enough to produce serious candidates against a disease the market ignores, then the list of forgotten diseases becomes a list of workable projects. Snakebite, Chagas disease and leishmaniasis have been waiting for decades.

Are we really going to build a tower of Babel?

I have written it since the first day of allelu.ai: the real question is not whether we should use artificial intelligence, but what we want to put it at the service of. This scorpion venom story puts that question on the table with unusual clarity.

The tower, here, would have a recognisable face: the record that matters more than the patient. The figure that circulates, the show of power, and in two months nobody asking what became of the 17 candidates, while the children who were stung go on waiting for a vial.

The city is the same technology, with one decision added. Deciding that diseases without a market deserve the computing power. Publishing the data instead of locking it away. Accepting review from teams who gain nothing if it works. And measuring success by the number of people treated, not the number of molecules generated.

Nothing is settled. For now, this project leans the right way: a useful target, open data, outside review, stated caution about what has not been shown. What is still missing is the commitment to what comes next, the part that costs a lot and makes no headlines. That no longer depends on the machine. AI is a revolution. Humanity remains our responsibility.

And you?

If you had to choose the next forgotten disease to hand to an artificial intelligence, which would you choose, and why that one?

If this article was useful to you, sign up for the Thursday newsletter: one article a week to understand AI and its ethical questions, without the jargon. And forward it to someone who ought to read it.

To go further on how double-edged these tools are, you can read AI, the fire of the 21st century and When AI gets out of its cage.

---

Reynald Naulleau is the author of L'abondance tuera (in French) and the founder of Allelu AI (allelu.ai), a space for understanding artificial intelligence and ethics, without the jargon.

Sources

• Anthropic, "How Claude designed protein binders against scorpion neurotoxins", October 2026
• Anthropic, full technical report, PDF
• The Decoder, independent write-up
• Summary of the encyclical Magnifica humanitas, Leo XIV, 15 May 2026 (in French)

Article written on 24 October 2026. The figures for stings and deaths are those cited by the source and remain approximate. Written with the help of AI, reviewed and approved by Reynald Naulleau.