# Claude Unshackled: AI Designs Experimental Antivenom > Source: > Published: 2026-10-20 13:16:00+00:00 *Based on Hunterbrook Media’s reporting, at the time of publication Hunterbrook Capital is both long and short securities related to AI in biotech. Positions may change at any time. This article is not investment advice or any recommendation. See full disclosures below.* It’s still only a laptop on a table, in a sense. But when you see a scientist at Anthropic deploy an advanced, unconstrained model, you see why there’s a chasm between the company’s [eschatological warnings](https://darioamodei.com/essay/the-adolescence-of-technology) and many people’s experience of AI. We aren’t actually using the same tools the same way. In that room at Anthropic, the unreleased Claude ran much faster. It was not limited by compute. The prompts were thoroughly engineered, including internal tactics I wasn’t allowed to record. And oh boy, if you’ve ever hit the frustrating biosecurity limits of Anthropic’s most advanced public model (Fable, which demotes bioscience work to [a less fearsome model](https://www.anthropic.com/claude/mythos)), it’s a striking experience to see for the first time a truly frontier model unleashed against a biological target. *Claude — your mission, should you choose to accept it: Design new molecules in a computer that we can then make in a lab as an experimental antivenom that binds scorpion toxins.* An initial prompt (dozens of pages longer than the above) and a few efficient follow-ups guided a campaign that selected 211 designs. But then we had to actually test the molecules. Preclinical labwork is where we hit the speed limit: physical science. Contracts and orders. Shipping. Contractors or subcontractors facing delayed materials. Slow results. Emails. So while patients, doctors, companies, and investors hope for the potential of AI in biotech, real-world barriers reify how hard it is to research new medicines. Weeks after the designs, we got the initial preclinical data. It’s promising, but mixed. A useful experiment on the frontier of AI-driven bioscience: its potential, its constraints. Then, once a new candidate molecule graduates preclinical studies, there are still multiple phases of clinical trials. Manufacturing. Safety risks. Regulatory gauntlets. Affordability and access issues. Insurance giants, insurance middlemen, [middlemen’s middlemen](https://hntrbrk.com/investigations/pbmgpo). So while software continues to devour the world, it must also gnaw on the gristly bones of medical development and actually getting medicines to patients across the world. The scorpion struck the guide as he stood up from the couch. Ian Breland had always been drawn to creatures that most try to avoid. He was kicking back after leading a snake-watching expedition near Victoria Falls in Zimbabwe. Then, the sting. The pain was immediate. “Just like a hot rod is being stuck through the bottom of my foot,” he told Hunterbrook. Pain seized his ankle. He could feel the venom travel up his leg. He lost all sensation below it: “You could have shot me with a BB gun.” He texted a photo of the scorpion to an expert friend. “Are you joking?” the friend replied. His crew rushed to help, but the nearest clinic was almost an hour away, and it didn’t have antivenom. Breland tried to text his parents. The phone screen wasn’t working, it was drenched: His fingertips had begun “sweating like a small faucet.” Blood drained from his head — “Whoa, guys” — and he blacked out. Scorpion venom is a mix of toxins that wreak havoc on the body. Scorpions sting [more than a million](https://pmc.ncbi.nlm.nih.gov/articles/PMC12636920/) people every year. Thousands die, explained Dr. Eduardo Hernández Muñoz, an epidemiologist with the Mexican public health system, who outlined the field to Hunterbrook. Fortunately, Breland wasn’t one of the fatalities. “These numbers we see from the WHO, of people dying, that’s hundreds and thousands of young kids," Breland said. Last October, a 5-year-old girl was stung at her preschool in Sonora, Mexico. She was taken to a local public hospital. But the hospital had already used up the antivenom delivered just days before, according to an [official statement](https://x.com/IMSSSONORA/status/1981539421204758844?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E1981539421204758844%7Ctwgr%5Ea75c9751cd316ae9c74c3ccd219407a66c154a52%7Ctwcon%5Es1_c10&ref_url=https%3A%2F%2Fwww.univision.com%2Flocal%2Farizona-ktvw%2Fnina-murio-alacran-hermosillo-ignacia-e-de-amante). The staff found more, 45 minutes after she’d arrived. But it was too late. She’d already suffered two cardiorespiratory arrests. Resuscitation failed. Soon after, a clinician pronounced her dead. ## **The venom and its cure** There are [many species](https://academic.oup.com/ije/article/53/3/dyae070/7675638) of scorpion, with varying types of venom and effects on their victims. As a result, antivenom is often a cocktail of [different molecules](https://www.dailymed.nlm.nih.gov/dailymed/drugInfo.cfm?setid=5cb65048-a30c-48e5-8bc8-897983d08068), infused together to bind different toxins. The aim: block the toxins from proving lethal. The sole [FDA-approved](https://news.arizona.edu/news/ua-scorpion-antivenom-collaboration-gains-fda-approval) scorpion antivenom is [Anascorp](https://www.anascorp-us.com/). There are variants across the world, using similar approaches: immunizing horses with venom, then harvesting venom-targeting antibodies from the horse’s immune system. Once purified, these proteins become the “secret sauce” of antivenom, according to [Jonathan Leeming](https://scorpions.co.za/), an expert on scorpion sting emergency response in Southern Africa. “The original way that they made the antivenom was like 100 years old. In that time, it was groundbreaking. It was like landing on the moon,” Leeming told Hunterbrook. “But in today's day and age, there's much better ways to produce a much better quality antivenom.” Developing antivenom for the most dangerous scorpion species in North America took [decades](https://biotecmov.ibt.unam.mx/numeros/32/2.html), starting in Mexico during the 1970s. Dr. Eugenio Vladimir Zavala Sánchez, a colleague of Hernández, told Hunterbrook that although the cocktail making up Anascorp has flaws, it’s currently among the most effective antivenoms in the world. And it needs to be: [About a quarter](https://www.scielo.org.mx/scielo.php?script=sci_arttext&pid=S0016-38132025000600691&lng=es&nrm=iso) of all scorpion envenomations are in Mexico. "Even though it's not perfect," Hernández echoed, "it works." “Antivenom remains the only treatment capable of directly neutralizing circulating venom toxins,” Dr. Jesús Chávez Guzmán, an emergency physician in Mexico, wrote to Hunterbrook. “These numbers we see from the WHO, of people dying, that’s hundreds and thousands of young kids." — Ian Breland After Mexican scientists spent 25 years developing the antivenom, it took their American collaborators another decade of clinical trials in the U.S. to get [FDA approval](https://news.arizona.edu/news/ua-scorpion-antivenom-collaboration-gains-fda-approval). While other scorpion antivenoms across the world are now available or in development, it is not a major focus of the pharmaceutical industry, which typically [prioritizes](https://www.efpia.eu/media/uj0popel/the-pharmaceutical-industry-in-figures-2026.pdf#page=4) [more lucrative markets](https://pharmaphorum.com/news/most-pharma-firms-doing-little-rd-for-poorer-patients). Much like Anascorp in the U.S., a single laboratory produces scorpion antivenom in Mexico, Zavala Sánchez told Hunterbrook. The same is true in Southern Africa, where Leeming said the main supplier ran out by October, with no new stock expected until the end of 2026. The day of his Hunterbrook interview, Leeming had already responded to two calls for scorpion stings. He said there is “no reason to die from any envenomation,” but “there’s always extenuating circumstances.” Alongside access and education: the quality of the antivenom. The antivenom could be more effective and cheaper — if developed without horses, according to Leeming. And even purified horse-derived antibody products can trigger allergic reactions, including anaphylaxis, so there’s a need for antivenom with fewer side effects. ## **Could AI design new experimental antivenoms — and without poisoning horses?** Advocates of AI hope it will [dramatically accelerate](https://darioamodei.com/essay/machines-of-loving-grace) the development of new medicines, especially for underserved patients. “Scorpion envenomation has historically received less research attention than many other major infectious and non-communicable health priorities,” wrote Chávez Guzmán, “despite the large number of cases in endemic countries.” The design of new molecules is a critical step. It’s also *relatively* [cheaper and faster](https://www.cbo.gov/publication/57126) than most of what follows in the development process. “AI is supercharging drug discovery,” said [Dr. Amy Abernethy](https://www.insitro.com/leadership/amy-abernethy/), formerly a [senior leader](https://medschool.duke.edu/news/2021-distinguished-alumna-award-amy-p-abernethy-md94-hs94-01-phd) of the FDA. “But there's still questions around capital, questions around safety, questions about the right experiments to do.” “It moves the choke point to clinical development,” she said. “It’s a brick wall at one point or another.” [Daphne Koller](https://en.wikipedia.org/wiki/Daphne_Koller), CEO of the techbio company Insitro — where Abernethy serves on the board — rhetorically posed the challenge: "I can make molecules all day long, but so what?" "We do not need more molecules against the same set of targets,” Koller told Hunterbrook in an interview with Citrini. She advocated instead for a deeper understanding of biology and better target selection. When briefed about the AI-driven experiment, she responded: "It's great that we have this capability, but we still need to know where to aim that … the failures by and large today are because we have no idea what to make molecules against.” And it’s not just about selecting better targets, either. “Experimental antivenoms can neutralise lethality,” [Dr. Naoual Oukkache](https://en.wikipedia.org/wiki/Naoual_Oukkache), research director of an antivenom lab at the Pasteur Institute of Morocco, wrote to Hunterbrook. However, “They must be followed by rigorous pharmaceutical development, standardised manufacturing, and clinical validation before the products can be used in humans.” That development process includes computer modeling (*in silico*), test tube experiments (*in vitro*), lab animal studies (*in vivo*), and most importantly, human clinical trials.  “Artificial intelligence and computational biology can accelerate and rationalise several stages of research,” wrote Oukkache. “Molecules designed or optimised using AI are unlikely to replace polyclonal antivenoms immediately. They could instead be developed as complementary or adjunctive treatments.” Many steps won’t imminently speed up due to AI. Cells must be tested. Monkeys must be bred. Patients must be identified. Molecules must be manufactured at increasing scale and quality. “We've always had a funnel, and that funnel has always been top heavy, and as the top gets wider, and we don't expand the middle, that becomes a tighter and tighter bottleneck,” said Anna Marie Wagner, CEO of biomedical company Transfyr, who detailed the current state of the field to Hunterbrook alongside her co-founder Renee Węgrzyn. The organizations that invent and advance new medicines — universities, biotech companies, venture incubators, and research institutes — must physically synthesize, manufacture, and test candidates. This often entails a [global network](https://www.fda.gov/international-programs/fda-globalization) of facilities or suppliers. Then developers need to recruit brave subjects to enroll in clinical trials of experimental molecules, convincing sick people to [take a risk](https://www.nhlbi.nih.gov/research/clinical-trials/safety-benefits-risks) on a drug that may not work, and may even harm. Then there’s grueling [regulatory work](https://www.fda.gov/patients/drug-development-process/step-4-fda-drug-review) that involves many experts completing thousands of pages of esoteric government filings. AI may help [find and enroll](https://www.nature.com/articles/s41467-024-53081-z) patients. It already makes fulfilling [regulatory requirements](https://arxiv.org/abs/2509.09738) easier. Even then, we — the patients, their families, doctors, medical organizations or companies, and society as a whole — all still need to wait for the clinical results of each experimental trial. Waiting for weeks. Months. Often, years. That’s the real bulk of the [time and cost](https://www.cbo.gov/publication/57126) that go into bringing a new medicine to market. Now, though, the design of new medicines is often becoming easier, faster, and cheaper with AI. The diffusion of new technology is uneven and often unfair. As the sci-fi author William Gibson [supposedly](https://quoteinvestigator.com/2012/01/24/future-has-arrived/) said: “The future is already here, it's just not evenly distributed.” In a bland conference room during a sweaty August afternoon on the second floor of Anthropic’s office in New York City, we got a glimpse of that future. ##### ***This final section recounts the experience of Hunterbrook’s CEO contributing to the AI-designed antivenom project. Prior to Hunterbrook, he worked in the biomedical sector.*** *This final section recounts the experience of Hunterbrook’s CEO contributing to the AI-designed antivenom project. Prior to Hunterbrook, he worked in the biomedical sector.* As an erstwhile scientist, I felt the quickening pulse of discovery our first day at Anthropic. But as early as our second day, as we finished the key computer step — and especially after our second month waiting on labwork — I recognized a forgotten, more quotidian sensation. The *boredom* of biotechnology. Even with virtually unlimited resources, bioscience often remains really damn slow. The beauty and importance of biotech can erode during that wait: for materials, data, traction. That torpor was one of the reasons I largely left the field, after a decade of my life, to build Hunterbrook. Waiting on data from this experiment at Anthropic, I remembered earlier years of impatience. I’d danced on the frontier of biotech since I was a teenager. I began with a high school project about CRISPR the year it debuted; in college, I used that gene editing tool to help discover a new medical target for treating heart disease, work I [co-published]() in the scientific journal *Cell*. I dropped out of Harvard to lead a startup working on an [experimental medicine](https://www.thedailybeast.com/the-biotech-startup-taking-on-pancreatic-cancer-by-throwing-out-the-rules) for pancreatic cancer in Silicon Valley, where I also [encountered Elizabeth Holmes](https://nbhorwitz.medium.com/if-evil-be-said-of-thee-5177162a1b) (I later informally advised on “The Dropout” TV series; and The Wall Street Journal’s exposé on Theranos became part of the inspiration for Hunterbrook’s investigative newsroom). Like most novel biotech, my first experimental medicine [failed](https://www.washingtonpost.com/national/health-science/heres-what-happened-when-i-tried-to-develop-a-new-drug-for-a-deadly-cancer/2019/01/11/697cf9f6-039f-11e9-9122-82e98f91ee6f_story.html). I returned to college to graduate with a degree in molecular biology. Then I joined what became one of the premier biotech firms. Within that firm’s startup incubator, I helped [architect and scale up](https://nbhorwitz.medium.com/life-and-death-at-a-distance-ddbd2f6268e8) new biotech companies from scratch. Before the earliest iterations of Claude or ChatGPT were publicly released, two of my ventures were dedicated to using AI for biomedical design. In 2019, I co-founded [Triana](https://endpoints.news/pfizer-allied-molecular-glue-startup-triana-collects-120m-series-b/), using AI to [help design](https://trianabio.com/press-release-4-6-2022) a new class of medicine called molecular glue. Seven years and over a hundred million dollars later, Triana’s first candidate entered the clinic this spring with [a Phase I/II trial to treat lung cancer](https://clinicaltrials.gov/study/NCT07491497?intr=tri-611&viewType=Card&rank=1). In 2021, I co-founded Superluminal, using AI to rapidly optimize therapeutic molecules. Nvidia invested in our [$120 million](https://cen.acs.org/business/finance/Lilly-Nvidia-back-biotech-firm/102/i29) financing, in part so Superluminal could buy Nvidia’s compute. Superluminal recently [announced](https://www.biocentury.com/article/660717/superluminal-s-60m-series-b-to-drive-first-clinical-study-finance-report?editionId=cmtnpi9c1nyw006o4k6yqj8ee) it raised tens of millions in new funding to enter the clinic this fall, five years after its creation. Triana and Superluminal are both [AI-enabled companies](https://trianabio.com/press-release-031926-1) led by veteran executives, [well-funded](https://www.superluminalrx.com/wp-content/uploads/Superluminal_Series_A_Financing_Press_Release-090924.pdf) by top firms. Yet many years later, they still don’t know if each medicine will work in people, let alone merit FDA approvals. Software has aided drug discovery for decades, as Hunterbrook reported in a feature on one of the field’s pioneers, [Schrödinger](https://hntrbrk.com/investigations/schrodinger). AI really burst into the biotech zeitgeist when Google DeepMind’s AlphaFold software began solving key challenges in protein folding, for which two of its leaders shared a Nobel prize with a professor known for computational protein design. The [“techbio”](https://www.drugdiscoverytrends.com/assessing-the-techbio-landscape-hype-or-substance/) sector boomed during the pandemic. It busted as valuations overextended and interest rates rose. It may be booming again, as evidenced by the rise in certain techbio stocks. Recent biotech sentiment has been partly catalyzed by the increased focus of Anthropic and [OpenAI](https://openaifoundation.org/news/public-data-for-health) on how AI may help with biomedical research, a more sympathetic narrative than displacing jobs or risking doom. Anthropic even [acquired](https://www.thestar.com.my/tech/tech-news/2026/09/18/exclusive-anthropic-quietly-sets-up-biology-lab-as-it-ramps-ai-drug-program) a biotech company and opened its own biomedical research lab. But what do those increased resources actually get you compared to what was possible before? After Claude designed our molecules, Anthropic and Citrini paid two outsourced labs — each a contract research organization (CRO) — to synthesize and test 211 designs. It started simply: Claude emailed a go-to Anthropic contractor with the molecules to synthesize. The contractor shared an invoice, a timeline, and next steps for testing the synthesized binders. Then, almost immediately — as so many discovery projects do — the timeline began to slip. *Here we go, I thought, recalling similar frustrations, often involving even simpler projects.* While synthesizing the binders was relatively straightforward, sourcing the scorpion toxins and certain other reagents to test the binders proved slower. We now have initial molecular binders for four of the five scorpion toxins we targeted; the fifth was a messier result. The experiments did not establish that these binders neutralize toxins or will work as an antivenom. And while the binders passed a test against one unrelated protein, broader testing flagged some as potentially sticky, meaning their selectivity is unconfirmed. Anthropic published the full results in its new [technical report] today, reviewed by independent academic experts. Although these early findings are promising, a new antivenom isn’t right around the corner. If this project were a full drug development program, it would still be at the earliest stages of an Odyssean expedition. Then again, no horses had to be poisoned with scorpion venom in the making of these molecules. Could a skilled team have done similar work using existing protein-design tools, without AI (and similarly horseless)? Yes. In this case, the AI did not break new ground specialists couldn’t have with available tools: Claude did not generate a binder that could meet the study’s most ambitious aim, meeting the criteria for a single molecule binding multiple toxins. However, based on my past work on comparable projects, without Claude the designs would have been slower, more expensive, and likely to entail a larger number of binders, due to a lower hit rate. That would also require more synthesis and more labwork. And other than one Anthropic scientist, none of the people directly involved were experts in drug discovery (or, in the case of the Citrini team leading the project, had any expertise in bioscience whatsoever). So the point is that — as hypothesized, but not yet proven — AI should make it much easier and more efficient to design molecules, including for non-experts like non-technical entrepreneurs or even [heroic parents](https://www.statnews.com/2024/12/09/pirovolakis-gene-therapy-rare-disease-spg50-lockard/). The number of new medicines starting the journey will likely multiply. Rather than a quantum leap, it feels more like another notable datapoint on an overall exponential trajectory, like cells proliferating on a petri dish, that biotech has been ascending since its inception, yielding miraculous discoveries across cancer and many other diseases. With this latest wave of AI, there will be faster progress. Though it’s hard to diagnose where one stands on a parabola of innovation, the second derivative of progress doesn’t feel vertical. Yet in biotech labs and AI labs across the world, the future increasingly presses against its limits. *“And what rough beast, its hour come round at last …”* **Nathaniel Horwitz** is the CEO of Hunterbrook Media and Hunterbrook Capital. He co-founded Hunterbrook after company-building and investing at a premier biotechnology and healthcare firm, where he was the youngest venture partner by over a decade. At that firm, Nathaniel incubated four companies, ranging from AI-designed medicines for cancer to cell therapies for autoimmune diseases. He has served on the board of directors at several companies and volunteers as executive chair for the health education nonprofit Mayday. He earned a degree in molecular biology from Harvard, where he co-authored biomedical research. He has published in several national outlets. His first job out of high school was reporting for his local newspaper. Nathaniel may contribute reporting or editorial input. **Laura Wadsten** is an investigative journalist specializing in healthcare. She began her career reporting on antitrust and healthcare as a Correspondent for The Capitol Forum, a premium financial publication. Laura was a Hodson Scholar and Editor-in-Chief of The News-Letter at Johns Hopkins University, where she earned a B.A. in Medicine, Science & the Humanities. **Vikas Kumar** joined Hunterbrook from The Capitol Forum, where he led the corporate investigations team for a decade as a senior editor. He was previously an attorney at Gordon Feinblatt, a trial attorney for the Department of Justice, and a law clerk for a federal judge. He has a J.D. from University of Virginia School of Law and a bachelor's from Emory University. Vikas is based in Maryland. **Dan DeLorenzo** is a creative director with 25 years reporting news through visuals. Since first joining a newsroom graphics department in 2001, he has built teams at Bloomberg News, Bridgewater Associates, and the United Nations, and published groundbreaking visual journalism at The Wall Street Journal, Associated Press, The New York Times, and Business Insider. A passion for the craft has landed him at the helm of newsroom teams, on the ground in humanitarian emergencies, and at the epicenter of the world’s largest hedge fund. He runs DGFX Studio, a creative agency serving top organizations in media, finance, and civil society with data visualization, cartography, and strategic visual intelligence. He moonlights as a professional sailor working toward a USCG captain’s license and is a certified Pilates instructor.