Anthropic has announced that its latest model, Claude Mythos, has not been publicly released. Instead, it is being made available to a highly restricted set of preview partners under a newly launched initiative called "Project Glasswing." While Claude Mythos is a general-purpose model, similar to Claude Opus 4.6, Anthropic asserts that its cybersecurity research capabilities are so potent that the entire software industry requires time to prepare for its implications.
The Mythos Preview has already identified thousands of high-severity vulnerabilities, affecting every major operating system and web browser. Given the rapid pace of AI progress, it is anticipated that such capabilities will soon proliferate, potentially extending beyond entities committed to their safe deployment.
Partners participating in Project Glasswing will gain access to Claude Mythos Preview to identify and remediate vulnerabilities or weaknesses within their foundational systems, which represent a significant portion of the world’s shared cyberattack surface. This work is expected to concentrate on tasks such as local vulnerability detection, black-box testing of binaries, securing endpoints, and penetration testing of systems.
Further technical insights are detailed in Anthropic’s Red Team blog post, "Assessing Claude Mythos Preview’s cybersecurity capabilities." For instance, Mythos Preview successfully crafted a web browser exploit that chained together four vulnerabilities, employing a complex JIT heap spray to bypass both renderer and OS sandboxes. It autonomously achieved local privilege escalation exploits on Linux and other operating systems by exploiting subtle race conditions and KASLR-bypasses. Furthermore, it autonomously developed a remote code execution exploit for FreeBSD’s NFS server, granting full root access to unauthenticated users by distributing a 20-gadget ROP chain across multiple packets.
A stark comparison with Claude 4.6 Opus highlights Mythos’s advanced capabilities. Internal evaluations showed that Opus 4.6 generally had a near-0% success rate at autonomous exploit development. Mythos Preview, however, operates at a profoundly different level. For example, while Opus 4.6 managed to turn vulnerabilities it found in Mozilla’s Firefox 147 JavaScript engine into JavaScript shell exploits only twice out of several hundred attempts, Mythos Preview, when benchmarked, developed 181 working exploits and achieved register control in 29 additional instances.
Anthropic's caution appears well-founded. Recently, credible security professionals have increasingly raised alarms regarding the enhanced capabilities of modern Large Language Models (LLMs) in vulnerability research. Greg Kroah-Hartman of the Linux kernel noted that months prior, they received "AI slop"—obviously wrong or low-quality AI-generated security reports—but the situation has since dramatically shifted, indicating a rise in sophisticated AI-driven vulnerability discoveries.