Anthropic's Claude Mythos Finds Thousands of Zero-Days in Major Software

Anthropic just revealed something that should make every software company pay attention — and possibly lose a little sleep. Claude Mythos Preview, the company's most capable model ever built, has autonomously discovered thousands of previously unknown security vulnerabilities across every major operating system and web browser.

The model won't be available to the public. Instead, Anthropic is channeling its capabilities through Project Glasswing, a coordinated initiative that gives roughly 50 organizations access to Mythos for defensive cybersecurity purposes. It's an unprecedented approach: building the most powerful AI model in the world, then deliberately restricting it to protect critical infrastructure.

What Mythos Actually Found

The numbers are striking. Working entirely autonomously — no human guidance, no hints about where to look — Claude Mythos Preview identified thousands of high-severity zero-day vulnerabilities (flaws that developers didn't know existed) across foundational software.

Among the most notable discoveries: a 27-year-old vulnerability in OpenBSD that could crash systems remotely, and a 16-year-old flaw in FFmpeg (the multimedia framework used by nearly every video application) that automated security tools had missed despite running over 5 million times. The model also uncovered multiple Linux kernel vulnerabilities enabling privilege escalation — the kind of flaws that could give attackers root access to servers.

On CyberGym's vulnerability reproduction benchmark, Mythos Preview scored 83.1%, compared to 66.6% for the prior-generation Claude Opus 4.6. That's not an incremental improvement — it's a generational leap in AI-driven security analysis.

Why Anthropic Won't Release It

Anthropic's decision to restrict Mythos is deliberate and reflects a growing tension in AI development. A model capable of finding zero-days autonomously is extraordinarily valuable for defense. It's also extraordinarily dangerous if misused.

The company stated plainly that it does not plan to make Claude Mythos Preview generally available. Instead, the goal is to develop safeguards that can eventually be deployed with future models. Anthropic plans to launch new security controls with an upcoming Claude Opus model, using it as a testing ground for protections that don't carry the same risk profile as Mythos Preview.

This approach — build the most powerful thing possible, then figure out how to deploy it safely — represents a different philosophy than the "ship fast, fix later" approach common in tech. Whether it will hold as competitive pressure mounts remains an open question.

Project Glasswing's Partners

The initiative launched with 12 founding partners: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks. More than 40 additional organizations responsible for maintaining critical software infrastructure also received access.

Anthropic is backing the effort with $100 million in usage credits and $4 million in direct donations to open-source security organizations. That financial commitment signals this isn't a PR exercise — it's a sustained investment in infrastructure security.

The partner list is notable for its breadth. Apple and Microsoft rarely appear on the same initiative, yet both recognized that AI-discovered vulnerabilities in their operating systems represent a shared threat. The Linux Foundation's inclusion ensures open-source software — which underpins much of the internet — gets the same level of protection.

The Bigger Picture

Project Glasswing arrives at a critical moment for cybersecurity. Nation-state actors are increasingly using AI to discover and exploit vulnerabilities. Anthropic's bet is that giving defenders access to frontier AI capabilities first creates an asymmetric advantage — the good guys find the bugs before the bad guys do.

The implications extend beyond bug-finding. If Mythos-class models can autonomously identify vulnerabilities that human researchers and traditional tools missed for decades, it suggests AI is approaching a threshold where it can meaningfully outperform human experts at certain security tasks.

For enterprises, this means the standard for software security is about to rise dramatically. Vulnerabilities that might have survived undetected for years in critical infrastructure could now be found in hours.

What Comes Next

The immediate focus is patching. Every vulnerability Mythos discovered through Project Glasswing is being disclosed through coordinated disclosure processes, giving affected software maintainers time to develop fixes before details become public.

Longer term, Anthropic has signaled it will incorporate lessons from the Mythos deployment into its commercial products. The safeguards being developed now could eventually allow enterprises to run similar security analyses using future Claude models — without the risks that come with unrestricted access to Mythos-class capabilities.

For the cybersecurity industry, the message is unmistakable: AI-powered vulnerability discovery is no longer theoretical. It's happening now, and the organizations that adapt fastest will have a significant edge.