Anthropic debuts Project Glasswing, leveraging its powerful Mythos model to reinforce software security – News

Anthropic PBC said today it’s releasing a preview of the most powerful frontier model it has ever developed, making it available to a small coterie of partners and cybersecurity researchers to help secure the world’s software.

The model, called Claude Mythos, is being released as part of a new cybersecurity initiative dubbed Project Glasswing, which will see more than 40 partners use it specifically for “defensive security work.” According to Anthropic, though Mythos was not originally trained for security purposes, it excels at proprietary and open-source software code for vulnerabilities.

The company said it’s not releasing Mythos to the public because it’s just “too powerful,” and therefore too risky to make such a move.

Claude Mythos was first revealed in March in a leak that was surfaced by Fortune. According to that report, the leaked details described Mythos as “larger and more intelligent” than Anthropic’s existing Claude Opus models, which are its most powerful publicly available offerings. It was initially designed to be a general-purpose model for Claude, designed to have exceptionally strong coding and reasoning skills that would enable it to perform tasks such as building AI agents and writing code.

Anthropic says caution is necessary because the “capabilities we’ve observed in Mythos Preview could reshape cybersecurity.” In the past few weeks, while testing Mythos, the company said it has identified “thousands of vulnerabilities” across websites and apps, including every major operating system and web browser in use today.

The partner organizations in Project Glasswing include Amazon.com Inc., Apple Inc., Broadcom Inc., Cisco Systems Inc., CrowdStrike Holdings Inc., the Linux Foundation, Microsoft Corp. and Palo Alto Networks Inc. In addition, access will be provided to about 40 other organizations that build or maintain “critical software infrastructure.”

The partners will share what they learn from using Mythos with the rest of the technology community, so everyone can benefit from it and develop more secure software, Anthropic said. To facilitate the partners’ research, Anthropic has committed $100 million in usage credits to Project Glasswing, so those partners won’t be required to pay the application programming interface fees for their security testing and research.

The company is also said to be having “ongoing discussions” with U.S. government officials about giving them access to Mythos, though it’s possible that those negotiations are complicated by the company’s ongoing legal battle with the White House. That’s because Anthropic was recently labeled as a “supply chain risk” for refusing to let the Pentagon use Claude for autonomous weapons targeting or mass surveillance.

Regarding Mythos’ prowess, Anthropic explained that it recently discovered a 16-year-old vulnerability in FFmpeg, which is used by hundreds of applications to encode and decode video. The bug was discovered in a line of code that had been scanned more than 5 million times by traditional security tools without ever catching it. What’s worse is that Mythos is also powerful enough to immediately develop a sophisticated exploit for the vulnerabilities it discovers, potentially allowing attackers to immediately take advantage and start doing damage.

But though Mythos can be exceedingly dangerous, it can also be used for good. Cisco Chief Security and Trust Officer Anthony Grieco said his team has been using the model to find and fix security vulnerabilities across both hardware and software “at a pace and scale previously impossible.” He said it represents a “profound shift and a clear signal that the old ways of hardening systems are no longer sufficient.”

Anthropic said its eventual goal is to make it so that Mythos-class models can be deployed at large scale by the public, but for that to happen, it needs to develop cybersecurity safeguards that detect and block its most dangerous outputs.

Mythos will be especially useful for software developers, if those safeguards can ever be built and verified. On the SWE-bench Verified benchmark that gauges AI models’ coding abilities, Mythos was able to solve 93.9% of all problems, a much higher score than Claude Opus 4.6’s 80.8% accuracy rate. Moreover, Mythos achieved 77.8% accuracy on SWE-bench Pro, which is a more challenging evaluation, compared to just 53.4% for Opus 4.6.

Image: Anthropic