Yesterday, Anthropic announced Claude Mythos Preview, a new general-purpose model that it says is exceptionally good at finding security vulnerabilities in code. In fact, the model is so good that Anthropic has decided not to release Mythos Preview to the general public. Instead, it’s being released to a select group of companies that control OSes and other critical software.

Anthropic found thousands of vulnerabilities across every major OS and web browser with Mythos Preview, but used these three examples to illustrate their severity:

• Mythos Preview found a 27-year-old vulnerability in OpenBSD—which has a reputation as one of the most security-hardened operating systems in the world and is used to run firewalls and other critical infrastructure. The vulnerability allowed an attacker to remotely crash any machine running the operating system just by connecting to it;
• It also discovered a 16-year-old vulnerability in FFmpeg—which is used by innumerable pieces of software to encode and decode video—in a line of code that automated testing tools had hit five million times without ever catching the problem;
• The model autonomously found and chained together several vulnerabilities in the Linux kernel—the software that runs most of the world’s servers—to allow an attacker to escalate from ordinary user access to complete control of the machine.

A lengthy Frontier Red Team report brings the receipts for security researchers with an in-depth look at what Mythos Preview uncovered and the step change that the new model represents over Opus 4.6:

For example, Opus 4.6 turned the vulnerabilities it had found in Mozilla’s Firefox 147 JavaScript engine—all patched in Firefox 148—into JavaScript shell exploits only two times out of several hundred attempts. We re-ran this experiment as a benchmark for Mythos Preview, which developed working exploits 181 times, and achieved register control on 29 more.

As part of a test, Mythos Preview also managed to escape its sandboxed environment, message the researcher conducing the test, and then, outside the parameters of the test, posted about the exploit online.

The idea behind Project Glasswing, whose participants include Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, is to give them a head start at securing their systems before similar models emerge and are exploited for cyberattacks. If Mythos Preview’s capabilities are as Anthropic makes them out to be, this seems like the right approach. However, I do worry that with time, it could lead to a two-tier Internet where big tech companies operate in relative security thanks to tools like Mythos Preview, while those without access are left to swim with the sharks.