Anthropic is keeping its Mythos Preview model away from the general public, distributing it only to carefully vetted tech and security firms due to concerns about its extraordinary ability to discover and weaponize software vulnerabilities.
The company announced the restricted rollout Tuesday, making clear that the model's capabilities are too dangerous to release widely without first developing controls. Logan Graham, head of Anthropic's frontier red team, described Mythos Preview as "extremely autonomous" with sophisticated reasoning skills that match those of an advanced security researcher.
The scale of the threat is staggering. Mythos Preview can identify tens of thousands of vulnerabilities across software systems, a stunning leap from Anthropic's previous flagship model, Opus 4.6, which detected roughly 500 zero-day exploits. More alarming: the new model doesn't just find flaws, it automatically writes functional exploits to leverage them.
A Weapon Finding Hidden Targets Everywhere
Testing revealed how indiscriminate the model's power truly is. Mythos Preview uncovered bugs in every major operating system and web browser, including some dormant for decades that escaped detection through repeated manual security audits. It successfully crafted working proof-of-concept exploits on the first attempt 83.1 percent of the time.
One test case exposed the real-world risk: Mythos Preview found multiple flaws in the Linux kernel, then autonomously chained them together to enable complete system takeover on any machine running Linux. That's potentially catastrophic, given Linux powers most of the world's servers.
In another instance, the model discovered a 27-year-old vulnerability in OpenBSD, a hardened open-source operating system used in firewalls, routers and high-security servers. The flaw would allow attackers to remotely crash any system running it.
Graham acknowledged a hard truth: other AI companies are likely just months away from releasing similarly powerful models. "The security industry needs to understand that these capabilities may come soon," he said. OpenAI and other major tech firms are already developing comparable tools.
Instead of broad release, Anthropic is distributing Mythos Preview to more than 40 organizations for defensive security work on their own systems and codebases. Twelve major companies, including Amazon Web Services, Apple, Microsoft, Google, Cisco and Nvidia, are participating in a new initiative called Project Glasswing. The company is backing the effort with up to $100 million in usage credits for participating firms and $4 million for open-source security organizations like the Linux Foundation and Apache Software Foundation.
Anthropic has also briefed the Cybersecurity and Infrastructure Security Agency and the Commerce Department on both the risks and potential benefits of the technology. A company official framed the defensive opportunity plainly: "There's a chance here to give a shot in the arm to defense and to keep pace with this long-standing trend where offense exploitation had an advantage."
The restricted approach marks a departure from how the company initially handled the model. Graham noted Anthropic never formally intended to make this particular version publicly available. "The feedback was overwhelmingly clear to us," he explained. "We then decided to launch it this way."
Anthropic's eventual goal is to develop safeguards that allow broader deployment of Mythos-class models for general use cases beyond cybersecurity. The company plans to test new protections first on its less-powerful Opus models, allowing refinement without exposing the highest-risk version.
The situation underscores how AI capabilities are outpacing defenses. Prior incidents have already shown the danger: Chinese state actors weaponized Anthropic's models to automate a spying campaign targeting 30 organizations, and cybercriminals have used AI to write malware and automate ransomware negotiations.
Comments