Anthropic’s Claude Mythos found a vulnerability in OpenBSD that had been hiding since 1999. Twenty-seven years of audits by some of the best security engineers alive, and a machine found what they could not in hours. Then it found thousands more across every major operating system, browser, and cryptography library in production. The company locked the model inside Project Glasswing, restricted access to 12 enterprise and defense partners, and reported a 72.4% autonomous exploit generation rate, meaning Mythos could weaponize nearly three out of four vulnerabilities it discovered without human help. Securing critical infrastructure first was the obvious move.

But Mythos demonstrated something larger than a cybersecurity tool. It demonstrated that artificial intelligence can now identify structural flaws in complex systems that human experts have stared at for decades and missed. Software is one such system. Biology is another. And the gap between what frontier AI could do for disease research and what it is currently doing is worth examining in detail.

Introducing Project Glasswing: an urgent initiative to help secure the world’s most critical software.

It’s powered by our newest frontier model, Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans.https://t.co/NQ7IfEtYk7
— Anthropic (@AnthropicAI) April 7, 2026

The Precedent Is Already Clinical

This is not a hypothetical. AI-driven drug discovery has already produced molecules that are in human trials, with timelines that would have seemed absurd five years ago.

The clearest example is Insilico Medicine’s rentosertib, the first drug candidate where both the target and the molecule were discovered entirely by artificial intelligence. Insilico’s system identified TNIK, a kinase involved in fibrotic signaling, as a novel target for idiopathic pulmonary fibrosis in 2019. It designed the molecule within a year. By February 2021, the compound was a preclinical candidate. Phase 2a results, published in a peer-reviewed clinical study in June 2025, showed something remarkable: across 71 patients at 21 sites, participants on 60 milligrams daily saw forced vital capacity improve by 98.4 milliliters, while the placebo group declined by 62.3 milliliters. The drug did not just slow the disease. It reversed lung function loss. Total elapsed time from target identification to Phase 2a readout: six years. Traditional drug development averages 12 to 15.

Insilico has since nominated 22 preclinical drug candidates between 2021 and 2024, averaging 12 to 18 months from project initiation to nomination. The industry benchmark for that stage is two and a half to four years. Ten of those programs have entered human clinical trials.

Recursion Pharmaceuticals, running a 504-GPU supercomputer called BioHive-2, reported in December 2024 that its AI-discovered CDK7 inhibitor REC-617 produced a confirmed partial response in a patient with platinum-resistant ovarian cancer who had failed four prior lines of therapy. The response lasted more than six months. Recursion’s pipeline now includes more than 30 disease models in discovery and nine programs in preclinical development, all originating from its computational platform. Its REC-1245 program, an RBM39 degrader, went from target identification to investigational new drug application in under 18 months with about 200 compounds synthesized along the way. That timeline used to take four to five years.

The speed matters because it changes the economics. If you can identify a promising target and design a viable molecule in 18 months instead of five years, the cost of early-stage failure drops dramatically. Clinical trials remain expensive, but the preclinical phase, where more than 90% of drug candidates historically fail, is precisely where AI-driven pattern recognition compresses the most time and waste.

Editorial illustration of an eye cross-section with biological data patterns

What Pattern Recognition Finds When It Looks at Biology

The analogy between Mythos scanning source code and AI scanning biological data is imperfect, and the imperfection matters more than most AI-optimist writing admits. Software vulnerabilities are deterministic. A buffer overflow exists or it does not. When Mythos identifies a flaw, it can generate a working exploit that proves the vulnerability is real. The output is binary: the system is broken, or it is not. Biology offers no such certainty. When an AI model identifies a potential drug target, it has produced a hypothesis, one that still requires years of wet-lab validation, animal models, toxicology screening, and phased clinical trials before anyone knows whether the finding is real. A computational biologist would rightly point out that “finding a vulnerability” in code and “finding a target” in a disease pathway are separated by an enormous gap in verifiability.

That objection is fair. It is also, I think, slightly beside the point. The bottleneck in neglected disease research is not the final stage of clinical validation. It is the first stage: identifying which targets are worth pursuing at all. Researchers studying rare diseases often spend years narrowing candidates through painstaking manual analysis of genomic data, protein interaction networks, and published literature, work that is labor-intensive, underfunded, and largely sequential. That is the stage where Mythos-class pattern recognition could matter most, not by replacing wet-lab verification but by compressing the search space that precedes it.

Consider what Mythos actually does at a technical level. It ingests millions of lines of interconnected logic, maps dependencies across modules that were written by different authors at different times, identifies where assumptions in one layer contradict constraints in another, and produces a working proof that the flaw is exploitable. The human genome contains 3.2 billion base pairs. Protein folding, gene expression, and metabolic pathways are, functionally, the same kind of interconnected dependency graph. A mutation in one gene can cascade through regulatory networks in ways that only manifest as disease decades later, the biological equivalent of a latent vulnerability triggered by a rare input. The reason these flaws go undiscovered is the same reason the OpenBSD bug survived 27 years: the system is too large and too interconnected for human review to catch everything.

When DeepMind released AlphaFold, it predicted the three-dimensional structures of more than 200 million proteins. Demis Hassabis and John Jumper won the 2024 Nobel Prize in Chemistry for that work. But the downstream applications are what matter here. AlphaFold’s predictions revealed the structure of ApoB100, a protein central to atherosclerosis and the leading cause of death globally, whose structure had eluded researchers for decades. It identified 353 previously unknown proteins in the malaria parasite, reducing the uncharacterized protein count by 25%. It elucidated the structure of huntingtin, the misfolded protein behind Huntington’s disease, giving researchers their first clear picture of how the toxic aggregation that destroys neurons actually works.

Isomorphic Labs, the DeepMind spinoff focused exclusively on drug discovery, has built proprietary models that outperform AlphaFold 3 at predicting binding affinity, the measure of how strongly a drug candidate interacts with its target. The company secured $45 million upfront from Eli Lilly with $1.7 billion in milestone payments, $37.5 million upfront from Novartis with $1.2 billion in milestones, and a $600 million funding round from Thrive Capital in March 2025. It is preparing to launch clinical trials for AI-designed drugs in oncology and immunology.

Then there is retinal imaging. Ophthalmologist Pearse Keane’s team demonstrated that a transformer model trained on retinal scans can detect diabetes, hypertension, kidney disease, liver disease, coronary artery disease, and neurodegenerative conditions from photographs of the eye alone. The model identified markers for Alzheimer’s disease seven years before symptom onset and Parkinson’s five years before. Eric Topol, director of the Scripps Research Translational Institute, described the capability at the National Institutes of Health Clinical Center in 2024: “Machine eyes will see things that humans will never see.”

Editorial illustration of global drug research funding disparity

Where the Capability Gap Meets the Funding Gap

All of the examples above share a common feature: they involve diseases where significant research capital was already flowing. Pulmonary fibrosis, cancer, atherosclerosis, Alzheimer’s. These are conditions with large patient populations in wealthy countries, conditions that pharmaceutical companies can monetize.

The diseases that would benefit most from AI-driven pattern recognition are the ones where research capital is thinnest. Of the more than 7,000 known rare diseases, only about 5% have any approved treatment, according to the United States Food and Drug Administration. Neglected tropical diseases like Chagas disease, leishmaniasis, and schistosomiasis affect more than one billion people worldwide, per the World Health Organization. Most receive less than $100 million annually in research and development funding. Roughly 65% of global infectious disease R&D investment goes to three conditions: HIV/AIDS, tuberculosis, and malaria. The remaining 18 diseases on the WHO’s neglected tropical disease list split what is left.

Gene therapy, which once promised to address rare genetic conditions at their source, saw venture capital investment collapse from $8.2 billion in 2021 to $1.4 billion in 2024, according to PitchBook. The biology did not get harder. The economics got less favorable.

To put the disparity in concrete terms: Recursion Pharmaceuticals operates a 504-GPU supercomputer and has raised over $1 billion to apply AI to oncology and metabolic disease. There is no equivalent effort for Chagas disease, which infects an estimated six million people across Latin America and kills more than 10,000 annually, according to the WHO. There is no BioHive-2 equivalent running genomic screens for schistosomiasis, which causes more than 200,000 deaths per year. Leishmaniasis, fatal if untreated, still relies on treatment protocols developed in the 1930s. The Drugs for Neglected Diseases initiative, one of the few organizations dedicated to this space, operates on an annual budget of about $70 million, less than what a single mid-stage pharmaceutical clinical trial costs.

The average cost to bring a single drug to market is $2.87 billion across its full lifecycle, per the Tufts Center for the Study of Drug Development. For a disease affecting 10,000 patients, that math does not work under any conventional pharmaceutical business model. But if AI can compress the timeline from 12 years to six, cut the failure rate by identifying better targets earlier, and reduce the cost of preclinical development by an order of magnitude, the calculus shifts. Insilico has already demonstrated that the first two are possible. The third is what a Mythos-class system, pointed at biological data instead of source code, could accelerate.

📢 AI-enabled drug discovery reaches clinical milestone https://t.co/0iPNKvcuRJ

Few AI-designed drug candidates have gone beyond in silico benchmarks. Now, a study in @NatureMedicine @biogerontology reports a successful phase 2a trial of rentosertib, an AI-discovered drug and… pic.twitter.com/5e5C4QbRl0
— Marinka Zitnik (@marinkazitnik) July 1, 2025

What Comes Next

The infrastructure to connect frontier AI to disease research is closer than most people realize, and further than it should be. The NIH’s Bridge2AI program is building standardized biomedical datasets designed specifically for machine learning analysis. The Department of Health and Human Services released its first department-wide AI strategy in late 2025, directing the CDC, FDA, NIH, and CMS to share data and code across a unified ecosystem. AlphaFold’s training data came from the Protein Data Bank, a public resource. Genomic sequences from the NIH’s National Center for Biotechnology Information are freely accessible. The WHO’s Global Health Observatory publishes epidemiological datasets for every neglected tropical disease on its list. The raw material for a biomedical Mythos already exists. What does not exist is a mechanism to connect it to frontier models.

Think about what that would require in practice. Dario Amodei, Anthropic’s chief executive, would need to decide that Glasswing’s second phase includes partners outside the technology and defense sectors. Monica Bertagnolli, the NIH director, would need to negotiate model access the way her agency already negotiates access to supercomputing clusters at Oak Ridge and Argonne national laboratories. The Drugs for Neglected Diseases initiative, led by Charles Mowbray, would need a frontier lab partner the way CrowdStrike and Cisco became cybersecurity partners: bringing domain expertise in tropical disease biology to a system built for finding what humans miss. None of these steps involve new technology. They involve institutional will from people with names and titles and the authority to say yes.

I keep returning to that OpenBSD vulnerability. Twenty-seven years. Thousands of expert reviews. And a machine found it in hours because it could process the entire codebase at a scale no human team could replicate. Somewhere in the genome of the Chagas parasite, or in the protein interactions that drive Huntington’s progression, or in the metabolic pathways behind one of the 6,650 rare diseases with no treatment, there are equivalent flaws waiting. The tools to find them exist. The question is when we decide to use them.

Editor’s note: This article was written by an AI agent in collaboration with Claude Opus 4.6.