Hacking DNA: What Anthropic's Mythos Model Means for Medicine

Anthropic just announced Claude Mythos Preview — an AI model so capable at finding software vulnerabilities that they won't release it publicly. 

Instead, they launched Project Glasswing with Apple, Google, Microsoft, and others, committing $100M to use the model defensively. In weeks, Mythos found thousands of zero-day vulnerabilities across every major operating system and browser — including one in OpenBSD hiding for 27 years.

But the cybersecurity headlines aren't the whole story. The same vulnerability chaining capability that links multiple software flaws into sophisticated exploits maps directly to how polygenic disease works — cascading gene interactions across multiple variants that we've never been able to trace. With models like Arc Institute's Evo 2 and DeepMind's AlphaGenome already decoding the genome, Mythos-class reasoning could change everything about how we understand and treat disease.

If we can hack code to break it, we can hack code to fix it — including the code that makes us sick.

Sources and full write-up: https://open.substack.com/pub/danmccoymd/p/hacking-dna-the-anthropic-story-nobodys

Transcript

SPEAKER_00 0:07

And Throbbit just announced a model they're calling Clawed Mythos Preview. It's so capable of finding software vulnerabilities that they won't release it publicly. Instead, they launched Project Glasswing, which is a consortium with Apple, Google, Microsoft, AWS, CrowdStrike, and others, and committed$100 million in credits to let these partners use it defensively. In the last few weeks, Mythos has found thousands of zero-day vulnerabilities across every major operating system and every major web browser. One bug in OpenBSD had been hiding for 27 years. And the model didn't just find individual flaws. It chained together three, four, sometimes five different vulnerabilities into sophisticated exploits, the kind of work that might take a human security researcher an entire day to do. Now everyone's going to talk about the cybersecurity angle, and they should, but here's where my mind goes. Code is code. Software is written in languages that follow rules, have structure, and contain bugs that interact in ways that nobody really anticipated. DNA, it's actually the same thing. Three billion base pairs of code written over millions of years, full of variants that individually seem harmless, but in combination cause devastating disease. That chaining capability where mythos links five small vulnerabilities into one sophisticated exploit, that's exactly how polygenic disease works. Most cancers, most autoimmune conditions, most cardiovascular disease, they aren't caused by one broken gene. They're caused by cascading interactions across multiple variants that we haven't really been able to trace. And this isn't just theoretical. Just last month, Arc Institute released EVO2, an AI model trained on 9 trillion DNA-based pairs that can read and generate genetic code across all domains of life. Keep mind's alpha genome is deciphering non-coding DNA. That's the 98% of the genome we used to call junk, and predicting how tiny changes there affect gene expression. Now imagine a mythos class model, one that can autonomously pursue long-range tasks, chain together subtle patterns, and find bugs hidden for decades. Now point that at the human genome. We've had the data for years. What we haven't had is a system that can reason across that much complexity autonomously. That's what just changed. The cybersecurity implications are real and serious, but the thing I can't stop thinking about is this. If we can hack the code to break it, we can hack the code to fix it, including the code that makes us sick. Sources in the full write up are in my sub stack and links in the description.