Anthropic announced Tuesday it will only provide preview access to its new Mythos model to a select group of tech and cybersecurity companies due to concerns about the AI's advanced hacking capabilities. The company is withholding public release until adequate safeguards can be implemented to control the model's most dangerous functions.
Mythos Preview demonstrates "extremely autonomous" behavior with sophisticated reasoning capabilities that match those of advanced security researchers, according to Logan Graham, head of Anthropic's frontier red team. The model represents a significant leap from the company's previous public release, Opus 4.6, which found approximately 500 zero-day vulnerabilities in open-source software.
In testing, Mythos Preview discovered vulnerabilities in every major operating system and web browser, including some believed to be decades old that had escaped detection in repeated human-led security assessments. The model successfully reproduced vulnerabilities and created proof-of-concept exploits on the first attempt in 83.1% of cases, finding tens of thousands of security flaws that would challenge even the most skilled bug hunters.
The model's capabilities extend to chaining multiple vulnerabilities together for maximum impact. In one demonstration, Mythos Preview identified several Linux kernel flaws and autonomously combined them to enable complete system takeover. In another case, it discovered a 27-year-old vulnerability in OpenBSD that could allow remote system crashes, despite the operating system's reputation for security hardening and use in critical infrastructure like firewalls and high-security servers.