Claude Mythos and the end of software

Claude Mythos preview is the first frontier model powerful enough that Anthropic has decided not to release it publicly. The model has been running internally since February 24th, accessible only to strategic partners like Project Glass Wing and Vertex AI on Google Cloud. This isn't a gradual capability increase — Mythos is to Opus what Opus was to Sonnet: a much larger, slower, more expensive model that crushes every benchmark thrown at it.

The security implications are what forced Anthropic's hand. In testing, Mythos autonomously discovered and exploited zero-day vulnerabilities in major operating systems and web browsers. It found a 27-year-old vulnerability in OpenBSD, one of the most security-hardened systems in the world, and a 16-year-old flaw in FFmpeg. Most alarmingly, it chained together several Linux kernel vulnerabilities to escalate from ordinary user to root access. These capabilities emerged not from explicit training on security, but as a byproduct of becoming exceptionally good at code.

Anthropoc is pricing Mythos at $25 per million input tokens and $125 per million output tokens — roughly 10× the cost of GPT-4. But cost is irrelevant when the model isn't for sale. The company has committed up to $100 million in usage credits and $4 million in direct donations to deploy Mythos defensively, partnering with tech giants and government agencies to patch infrastructure before adversaries catch up.

The most dangerous exploits don't come from people who deeply understand security in isolation. They come from researchers who combine security knowledge with deep expertise in obscure domains — font rendering libraries, Unicode text shaping, browser memory layouts. Security expert Thomas wrote that vulnerabilities hide not in obvious places like password storage, but by «following inputs across the circulatory system of a program, starting from whatever weird pores and sphincters that program happens to take user data from.» Elite researchers had to learn the anatomy of these systems because that knowledge unlocked high-value targets like browsers.

Humanity was shielded from catastrophic exploits not only by sound countermeasures, but by a scarcity of elite attention. Only a small number of people possessed both top-tier security skills and the auxiliary knowledge required to chain vulnerabilities. Mythos changes this. The model scores roughly 8 out of 10 in security capability — there are humans who are better. But those humans are 5 or 6 out of 10 in most other software domains. Mythos is 9 out of 10 or better in every category. It has the breadth of knowledge to combine obscure system details with exploitation techniques in ways that would take a human researcher months or years to discover.

This is why the «window between a vulnerability being discovered and being exploited by an adversary has collapsed,» in Crowdstrike's words. What once took months now happens in minutes with AI. And critically, Anthropic didn't train Mythos to be good at hacking — they trained it to be good at code, and exploitation emerged as a side effect.