Claude's Scorpion Toxin Binders: 17 of 211 Designs Bound, and None Has Been Tested Against Venom

Anthropic said Oct. 20 that an internal Claude research model designed protein binders against five scorpion toxins, and that 17 of the 211 designs sent for lab testing met its binding criteria. The company's research post says it has not shown that any of them neutralizes venom.

The work was done with Citrini Research. The model ran mostly on its own for 64 hours, guided by one initial prompt and four follow-ups, Anthropic said.

What Claude generated and what the labs tested

Claude used 12 design tools to produce 318,484 designs, filtered out toxin-like or already known proteins, and co-folded 553,052 design-toxin pairs, a computer prediction of how two proteins fit together. It picked 180 designs. Anthropic added 31 more, for 211. The campaign used about 10,000 GPU hours.

The targets were Cn2 (Mexico), Ts1 (Brazil), AaH I and AaH II (North Africa) and LqhIII (Middle East), each at the site where the toxin attaches to a sodium channel. Adaptyv Bio and Twist Bioscience tested the designs with surface plasmon resonance, an instrument that measures how tightly two molecules stick. LqhIII was tested only at Adaptyv.

A design counted only if it hit at both labs in every replicate, bound with a dissociation constant (KD) of 1 micromolar or tighter, and gave no signal against CLEC12A, an unrelated human protein.

1. Cn27 designs
2. AaH I4 designs
3. AaH II4 designs
4. LqhIII2 designs
5. Ts10 designs

Source: Anthropic research post, Oct. 20, 2026; Ts1 count from Kingy.ai, accessed 2026-10-21

Where the 8 percent comes from

Seventeen of 211 is about 8 percent. Designs aimed at a single toxin did better: 11 of 60, an 18 percent hit rate that Anthropic's authors call typical for protein design. No design built to hit several toxins met the criteria for more than one. Kingy.ai counts 151 of those multitoxin designs and reports that none bound Ts1.

Cn2-B1 and Cn2-B2 were the strongest, with single-digit nanomolar affinities, Anthropic said. Competition experiments suggest they bind the same site as the LR antibody, which has been reported to neutralize Cn2 in mice.

What the authors say they have not shown

Binding was measured against recombinant toxins, not toxin purified from venom, and no design was tested for neutralization in cells or protection in animals. The authors wrote: "We have not confirmed whether these binders neutralize scorpion venom in a cellular environment." They do not claim an antivenom.

Kingy.ai, which did not test the proteins, adds three caveats from the technical report. Two binders also stuck to ovalbumin in an extra assay. The final computational ranking did not reliably pick out binders within a design set. And there was no matched expert-team baseline, so Claude's contribution cannot be separated from its tools.

Susana Vázquez Torres, a postdoctoral researcher in the Structural Biology Programme at the Spanish National Cancer Research Centre (CNIO) in Madrid, said in the post that binding "is no longer the main bottleneck" for targets like these. Timothy P. Jenkins, an associate professor at the Technical University of Denmark, called the work a sign of an AI-driven shift in therapeutic discovery for neglected envenomings. Both comments appear on Anthropic's own page.

The sting numbers and the data release

Anthropic cites over 1.2 million scorpion stings and more than 3,250 deaths a year, mostly among children with limited medical access. Kingy.ai notes a 2025 BMJ Global Health commentary that cites over 1.5 million cases.

All 211 designs and their data are on Hugging Face, Anthropic said. The post contains no biosecurity section.

The claim is narrower than most lab science announcements. Anthropic earlier said Claude cut quantum circuit gates 17x on one task, and OpenAI's quasi-Riemann claim rests on a 49-page paper no journal has checked. Here the authors name their own gap.

The authors list venom-purified binding, patch-clamp testing and mouse studies as the next steps, and give no dates for any of them.