Previously, foreign media FORTUNE had reported on this Anthropic Testing a model called “Claude Mythos” that claims to be the most powerful currently AI This model is reportedly even more powerful than the widely acclaimed Claude Opus 4.6, leaving many users eagerly anticipating it.
Just earlier, Anthropic finally officially unveiled this brand new AI model: “Claude Mythos Preview.” As previously revealed, this model is really powerful, especially in the cybersecurity field—it can not only find security vulnerabilities that have existed in operating systems and browsers for years, but even automatically generate attack programs to verify whether the vulnerabilities can actually be exploited.
Claude Mythos Preview Officially Launches as AI Model Built for Cybersecurity
Claude Mythos Preview is Anthropic’s most powerful AI model to date. It’s essentially a general-purpose large language model and wasn’t specifically trained for cybersecurity. However, due to major breakthroughs in code understanding and reasoning capabilities, it performs exceptionally well in cybersecurity applications.
According to Anthropic’s testing, Mythos Preview discovered thousands of high-severity security vulnerabilities within just a few weeks, with coverage spanning almost all major operating systems and major browsers. What’s even more staggering is that some of these vulnerabilities had existed for over a decade or even two decades without ever being discovered by humans.
Here are a few specific examples:
- A 27-year-old OpenBSD vulnerability allows attackers to remotely cripple systems
- Researchers discover a memory corruption vulnerability in FFmpeg that has existed for over 16 years
- A remote code execution vulnerability has been discovered in FreeBSD’s NFS server, allowing anyone to gain root privileges without authorization.
- Found multiple Linux kernel vulnerabilities and can chain them together to perform privilege escalation attacks
In the security baseline test CyberGym, Mythos Preview achieved an 83.1% vulnerability reproduction success rate, compared to only 66.6% for Anthropic’s own general-purpose flagship model Claude Opus 4.6.
In another OSS-Fuzz test, Mythos Preview successfully converted 181 Firefox vulnerabilities into working exploits, while Opus 4.6 only completed 2, making the gap between the two enormous.
Because Mythos Preview’s cybersecurity capabilities are so powerful, Anthropic has stated that the model will not be released to the public. If such capabilities fell into the wrong hands, the consequences could be extremely dangerous.
However, Anthropic isn’t completely withholding it either—instead, they’re promoting the “Project Glasswing” initiative, applying Mythos Preview’s capabilities defensively to help critical global software and infrastructure proactively identify and patch vulnerabilities, getting ahead before similar capabilities are developed by others.
Project Glasswing’s first wave of partners includes Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, Microsoft, and NVIDIA.
Anthropic also revealed that after the research preview phase ends, Mythos Preview’s API pricing will be $25 per million input tokens and $125 per million output tokens, and will be available through platforms including the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.
Source: KOCPC Chinese