Artificial intelligence firm Anthropic has unveiled its most advanced model yet, Claude Mythos Preview, but the company is holding back a full public release due to serious cybersecurity concerns.

According to Anthropic, the new model surpasses its previous Claude Opus 4.6 and even rivals leading systems like Gemini 3.1 on several benchmarks. However, what sets Mythos apart is its ability to detect critical vulnerabilities in software systems at a scale far beyond human capability.

Why the Model Is Considered Powerful
Claude Mythos Preview is designed to identify thousands of zero-day vulnerabilities—hidden flaws in software that developers are unaware of. These vulnerabilities can be highly dangerous, as they may allow attackers to exploit systems before fixes are available.

The company claims the model has already identified weaknesses in widely used platforms, including web browsers, operating systems, and even the Linux kernel, which powers a majority of global servers.

Collaboration with Tech Companies
Anthropic is currently working with over 40 technology companies to responsibly test the model and fix vulnerabilities before they can be exploited. This controlled rollout highlights the potential risks associated with such a powerful AI system.

Serious Security Implications
In one notable instance, the model reportedly uncovered a 27-year-old vulnerability in OpenBSD, an operating system known for its strong security. Some of the detected flaws could potentially allow attackers to gain full control over affected systems.

Why It’s Not Fully Released Yet
Due to the potential misuse of such capabilities, Anthropic has decided against an immediate public launch. The company is prioritizing safety, ensuring that identified vulnerabilities are addressed before wider access is granted.