Claude Mythos Preview, Anthropic’s latest artificial intelligence model, has identified thousands of high-severity vulnerabilities across every major operating system and web browser, the company said in a blog post published April 8, 2026. The disclosure has prompted Anthropic to organize a coordinated industry response before the model’s capabilities become more widely available.
Rather than releasing the model publicly, Anthropic launched “Project Glasswing,” a controlled program that gives select partners access to Mythos’s bug-finding capabilities. The goal is to patch critical vulnerabilities before other AI labs deploy models with similar power but fewer safeguards.
Project Glasswing Partners and Scope
Anthropic said it is working with Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks under the Project Glasswing umbrella. Access has also been extended to more than 40 additional organizations that build or maintain critical software infrastructure. In parallel, Anthropic said it is sharing information with the United States government about the model’s potential for offensive and defensive use in cybersecurity and its national security implications.
Claude Mythos Preview Exploit Capabilities
The model’s capabilities go well beyond identifying bugs. Anthropic’s researchers said Mythos wrote a web browser exploit that chained four separate vulnerabilities together, producing a complex JIT heap spray that escaped both renderer and operating system sandboxes. That level of exploit chaining has historically required highly skilled human security researchers.
In quantitative terms, Mythos converted 72.4% of identified vulnerabilities into working exploits within Firefox’s JavaScript shell. It achieved register control in a further 11.6% of attempted attacks. By comparison, earlier Claude models, specifically Sonnet 4.6 and Opus 4.6, each produced only a single crash at tier 3 severity across roughly 7,000 entry points in open-source repositories. Mythos reached 595 crashes at tiers 1 and 2, added crashes at tiers 3 and 4, and achieved full control flow hijack on ten separate, fully patched targets at tier 5.
Real-World Vulnerabilities Uncovered
Anthropic provided several concrete examples of bugs the model found. One was a 27-year-old vulnerability in OpenBSD that would have allowed an attacker to crash a system simply by connecting to it. Another was a 16-year-old flaw in the FFmpeg library, which Anthropic said was “hit five million times by automated testing tools without ever catching the problem.” A third example involved an exploit chain in the Linux kernel that could grant an attacker root access to the host system.
“We regularly run our models against roughly a thousand open source repositories from the OSS-Fuzz corpus, and grade the worst crash they can produce on a five-tier ladder of increasing severity, ranging from basic crashes (tier 1) to complete control flow hijack (tier 5).”
Anthropic Frontier Red Team
Disclosure Status and Future Access
Despite the volume of issues found, Anthropic said fewer than 1% of the potential bugs uncovered have been fully patched so far. The company said it is conducting responsible disclosure for all identified vulnerabilities, though the sheer number of findings makes rapid remediation difficult. Anthropic stated it will not make Claude Mythos Preview available for general use, and is instead documenting its behavior through the model’s system card.
In the longer term, Anthropic said the restricted partner program is intended to help companies and institutions prepare for a future where models of this capability class become more common. The lab’s Frontier Red Team described the risk of an unrestricted release as potentially equivalent to placing advanced exploit tools in the hands of unskilled attackers. Anthropic said it is also sharing findings with the U.S. government to inform national security planning around AI-enabled cyber operations.
