Just weeks after the leak of its Claude source code, Anthropic launched an AI-driven vulnerability taskforce backed by the likes of Cisco and Nvidia, with a promise of bolstering U.S. cybersecurity infrastructure against its enemies.

Project Glasswing centers around Anthropic’s Claude Mythos Preview, a new large language model (LLM) claimed to far surpass currently available models to identify and prioritize remediation of vulnerabilities. The venture was launched with some all-star backing, and even a Netflix-style video where talking heads are filmed as they sit down in the intro for their interviews.

Participants on the project include security leads from Amazon Web Services (AWS), Broadcom, Cisco, Google, Microsoft, Nvidia, and Palo Alto Networks. Cisco’ Chief Security and Trust Officer Anthony Grieco claimed foundational work alongside Anthropic allowed it to “identify and fix security vulnerabilities across hardware and software at a pace and scale previously impossible.”

Anthropic specified Mythos Preview had found “thousands” of high-severity vulnerabilities across every major operating system and web browser without human steering. These included zero-day vulnerabilities that have laid undiscovered for up to 30 years, as affecting FFmpeg and OpenBSD, respectively.

The AI giant claimed today’s frontier AI models are already competitive with the “best humans” at finding and exploiting vulnerabilities. This possibly opens up critical software to exploitation while compounding the estimated $500 billion fallout of cybercrime worldwide, according to Gov AI research cited by Anthropic.

Mythos Preview is presented as the solution, a general-purpose LLM that’s seemingly Anthropic’s most powerful model yet when it comes to software engineering, with a particular knack for computer security tasks. Unverified Anthropic data comparing Mythos to its current mainline Claude Opus 4.6 model showed the former could reproduce 83.3% of vulnerabilities compared to 66.6% from the latter on the same set.

Anthropic said it does not currently plan to make Mythos publicly available, with its use limited to only its Glasswing partners as well as a group of more than 40 additional and unnamed organizations.

A bad quarter for Anthropic

Project Glasswing comes after a pressing few weeks for Anthropic. The San Francisco-based firm recently saw a leak of the full source code behind its flagship Claude Code agent. The incident was attributed to human error in the release packaging process as opposed to a security breach, a reminder perhaps that all the AI in the world can’t fix lapses in operational processes, even if it's very good at sorting systemic software risk.

A deep dive on Mythos’ cybersecurity abilities adds that humans are still in the loop: After a bug is found, a final Mythos agent verifies the vulnerability with a triage and human validation process preceding any public disclosure.

Prior to the leak, Anthropic became embroiled in a legal battle with American authorities after President Donald Trump ordered every U.S. government agency to blacklist all of Anthropic’s AI systems.

The decree came after Anthropic declined to change its terms of service to allow mass domestic surveillance and support fully autonomous weapons.

Perhaps in response to this admonishment, Anthropic CEO Dario Amodei comments in Project Glasswing’s launch video that the firm has spoken to officials across the U.S. government about Claude Mythos Preview and its offensive and defensive capabilities, while “offering to work with them and collaborate to assess the risks of these models and to help defend against the risks of these models.”

This adds to comments in Anthropic’s press release for Project Glasswing calling out AI-driven cybersecurity attacks that may “empower adversaries of the United States and its allies,” adding “the emergence of these cyber capabilities is another reason why the U.S. and its allies must maintain a decisive lead in AI technology.”