Josh Kale
@JoshKale
Anthropic built an AI model so powerful they refuse to release itโฆ Claude Mythos has arrived. This is insane.
Here's everything it can do:
> Cracked flaws in TLS, AES-GCM, and SSH. The encryption protecting your bank, your messages, your data.
> Escaped a secure sandbox via a multi-step exploit, then emailed the researcher running the test. He found out while eating a sandwich in a park.
> Previous best model turned Firefox bugs into exploits 2 times. Mythos: 181. Same bugs. 90ร improvement.
> Previous models achieved 0 full system takeovers across 7,000 targets. Mythos achieved 10.
> 100% solve rate on the Cybench cybersecurity benchmark. Every challenge. Every trial.
> Discovered a 27-year-old bug the entire security industry missed. Cost: $50.
> Built a full attack granting root server access from anywhere on the internet. No human helped. Under $1,000.
> Found thousands of security flaws in every major operating system and browser on earth. Autonomously.
> First AI to solve a corporate network attack simulation end-to-end. A task estimated at 10+ hours for a human expert.
> Engineers with zero security training asked it to find exploits before bed. Working code by morning.
> 93.9% on SWE-bench Verified vs 80.8%. 59% on SWE-bench Multimodal vs 27.1%. More than double. 64.7% on Humanity's Last Exam vs 53.1%.
> Anthropic's own staff report a 4ร productivity multiplier. It's the most capable model ever made. Anthropic also says it poses the greatest alignment risk they've ever seen.
> None of this was specifically trained. It just emerged.
Anthropic called in Apple, Google, Microsoft, Amazon, NVIDIA, and 7 others to help. Nobody in the public is allowed access to this. $100M committed. They're racing to patch the world before this proliferates.
https://x.com/JoshKale/status/2041614953459724776?s=20