Anthropic and Citrini Research Tried to Use Claude to Create Scorpion Antivenom. Experts Say They Fell Well Short.
Claude generated nearly 320,000 protein candidates in 64 hours, narrowing them to 211 designs; tests found 17 binders against four of five scorpion toxins, but none targeted more than one toxin. Experts called the work preliminary.
An investment research firm teamed up with Anthropic in 2026 to test whether an unreleased version of the Claude AI model could design a scorpion antivenom from scratch. The result: some early-stage protein candidates, but nothing close to an actual drug, according to experts who reviewed the work.
Citrini Research, a London-based investment firm, initiated the collaboration after hitting Claude’s biosecurity guardrails while researching drug candidates. After cold-contacting AI lab employees, Citrini found a willing partner in Anthropic, which agreed to remove restrictions, provide access to an advanced unreleased model, and cover a compute bill of nearly $50,000. The work was conducted from Anthropic’s offices. An additional $150,000 was spent sending protein designs to two external labs for testing.
Using a 35-page prompt and roughly 300 AI subagents, Claude generated nearly 320,000 protein candidates over 64 hours, eventually narrowing the list to 211 designs intended to bind to scorpion venom toxins. When tested by Swiss biotech Adaptyv Bio and US-based Twist Biosciences, 17 proteins successfully bound to four of the five target toxins — but none bound to more than one toxin at a time.
Experts told The Verge the results were preliminary at best. Andreas Hougaard Laustsen-Kiel, a professor of antibody technologies at the Technical University of Denmark, said finding a good binder against scorpion toxins is “easy-ish,” while “the hard part is to translate it into a medicine.” Two colleagues at the same institution called the findings “far from a medicine and hyped,” noting that safety, stability, and real-world efficacy remain entirely untested. Andreas Bender, a professor of machine learning in medicine at Khalifa University, added that designing proteins that bind to targets is “kind of” a solved problem even without AI.
Citrini, which acknowledged that “none of us could claim to know anything about AI as it related to biology,” nonetheless declared the experiment a qualified success and used it to pitch investment opportunities in AI-enabled drug discovery to paying subscribers. Anthropic, which is approaching a widely anticipated IPO and has stated interests in drug development and life sciences, did not respond to a request for comment.