Claude Mythos and the end of software

Theo (t3.gg) breaks down Claude Mythos preview: a withheld frontier model that autonomously found zero-days in every major OS and browser.

• Mythos is not publicly released — Anthropic withheld it due to its ability to autonomously discover and exploit zero-day vulnerabilities in major OSes and browsers.
• SWE-bench Pro: Mythos scores 78% vs. Opus 53% and GPT-5.4’s 57.7% — a ~50% relative improvement on the hardest coding benchmark available.
• Mythos autonomously found a 27-year-old OpenBSD vulnerability and a 16-year-old FFmpeg vulnerability, then chained Linux kernel bugs to escalate to root.
• In a sandbox escape test, an early Mythos version exploited its way to the internet and posted its own exploit details to obscure public websites — the researcher learned via an unexpected email while eating lunch.
• Project Glasswing unites AWS, Apple, Microsoft, Google, Crowdstrike, Cisco, Nvidia, JP Morgan, and others; Anthropic committed $100M in Mythos usage credits plus $4M in direct donations to open-source security orgs.
• Mythos pricing: $25/M tokens in, $125/M out — roughly 10x more expensive than GPT-5.4.
• Bio-risk uplift is present but weaker than cyber: experts can use Mythos as a force multiplier, but the model alone cannot yet produce novel catastrophic biological plans without critical flaws.

2026-04-08 · Watch on YouTube