A360
Key Takeaways✦ Atlas AI
01
Anthropic's new Mythos AI model is being deployed in a limited release with partners like Apple and Microsoft for defensive cybersecurity.
02
The AI has reportedly already uncovered thousands of software vulnerabilities, including critical ones that are one to two decades old.
03
Project Glasswing marks a significant collaboration among tech competitors aiming to secure shared software ecosystems using advanced AI.
Atlas AI
Anthropic said its frontier AI model, Mythos, is being rolled out to a select group of major technology and security organizations to help identify serious software weaknesses before they are exploited. The company announced on Tuesday that the effort is being organized under a new, restricted-access program called Project Glasswing, designed specifically for defensive security use.
According to the announcement, 12 partner organizations will deploy Mythos with the aim of scanning both proprietary systems and open-source code for critical vulnerabilities. Anthropic framed the initiative as a proactive attempt to strengthen software security by using advanced AI to surface issues that traditional review processes can miss, particularly across large and complex codebases.
Project Glasswing brings together companies that typically compete with one another, including Amazon, Apple, Microsoft, and Google. The partner list also includes Broadcom, Cisco, CrowdStrike, the Linux Foundation, and Palo Alto Networks, which Anthropic described as a coordinated effort to reinforce digital infrastructure.
Under the program, partners are expected to use Mythos and then share relevant findings with the broader technology industry, rather than keeping results confined to a single organization.
Anthropic said the initial group is limited to 12 organizations, but preview access is planned to expand to 40 organizations in total. The company also said Mythos will not be released for general availability at this time, emphasizing that access is being tightly controlled due to the model’s specialized and sensitive use case.
On the model itself, Anthropic described Mythos as a general-purpose frontier system with particularly strong agentic coding and reasoning capabilities. A previously leaked memo referred to Mythos as one of the company’s “most powerful” models to date and positioned it for highly complex tasks. Anthropic also said that while Mythos was not trained specifically for cybersecurity, it has shown notable effectiveness when applied to vulnerability discovery.
In testing conducted over the past few weeks, Anthropic said Mythos has already found “thousands of zero-day vulnerabilities.” The company said many of the issues were critical, and that some had remained undetected in codebases for one to two decades. Those claims, if borne out through partner validation and remediation, would underscore how AI-based approaches can surface long-standing risks embedded in widely used software.
Anthropic positioned Project Glasswing as a test of how advanced AI can be used responsibly to protect the software that underpins global systems. The company said outcomes will depend on whether partners can act on the model’s outputs and communicate findings effectively to the wider community.
For markets and governments that rely on resilient digital infrastructure, the initiative highlights both the potential benefits of AI-assisted defense and the uncertainty around how quickly discoveries can be verified, fixed, and shared at scale.
