With Claude Mythos Preview, Zscaler integrates offensive AI into its development cycle

Zscaler has expanded its Project Glasswing initiative to incorporate Claude Mythos Preview into its secure development cycle, leading to a radical architectural conclusion: in an environment where vulnerabilities are discovered at machine speed, reactive patching is no longer sufficient. The solution is structural—reducing the attack surface to the point where adversarial AI has nothing to exploit.

By integrating Mythos Preview into its secure software development lifecycle, Zscaler aims to identify vulnerabilities within its application stack and Zero Trust Exchange platform before sharing the findings with the broader community. This approach underscores a belief that Zscaler has held for years and which has become imperative with the introduction of Mythos. “When the cloud emerged, the industry claimed that the traditional perimeter would hold. It did not. When mobile and SaaS made their debut, the industry asserted that VPNs would adapt. They did not. Every twenty to thirty years, architecture must change, and the companies that adapt will thrive in the following decade,” the company stated in its launch blog. Integrating Mythos into Project Glasswing is a practical embodiment of this belief.

Zscaler leverages a substantial detection infrastructure. Its platform processes over 500 billion transactions daily, with 40% of the Global 2000 as clients. This vast volume of signals forms the foundation of its behavioral detection capabilities. On this basis, Zscaler plans to construct a response to threats augmented by AI, focusing on making its architecture less vulnerable from the outset rather than merely accelerating the patching cycle.

Mythos Preview in Zscaler’s Development Cycle

Zscaler’s integration into Project Glasswing encompasses two key areas. The first is defensive, focusing on the company’s own assets. Mythos Preview is deployed early in the development cycle to identify vulnerabilities in the Zero Trust Exchange code before it goes into production. The results are shared with the program community, adhering to the collective sharing principle that underpins Project Glasswing. The second area is commercial. Zscaler is incorporating Claude Opus 4.7, the recently launched production model from Anthropic, into its AI Red Teaming and Agentic SecOps offerings. These products aim to detect and contain threats posed by autonomous AI agents capable of accessing data and interacting with other systems without direct human oversight.

This dual deployment illustrates the strategy Anthropic has built around the Mythos family. Mythos Preview remains restricted to high-sensitivity offensive-defensive work. Opus 4.7, designed with enhanced cyber safeguards and a deliberately reduced offensive capability compared to Mythos, is launched for legitimate SecOps use. For CISOs evaluating security automation tools, this distinction between the two models is not merely technical; it delineates what can be deployed in a production environment within a SOC versus what remains in the realm of controlled research.

A Two-Tier Cyber Defense

The gradual expansion of Project Glasswing creates a defensive asymmetry that disadvantages organizations outside the coalition. This results in a sort of two-tier cyber defense. Members detect and rectify critical vulnerabilities in their own assets before these flaws become known to the wider market. For European companies relying on a Zero Trust Exchange platform audited by Mythos, this delta translates to a higher level of hardening compared to non-member platforms. Currently, no equivalent European initiative offers similar access to critical infrastructure operators across the continent.

However, the implications of the Cloud Act remain stringent for all coalition members. Integrating Mythos Preview or Opus 4.7 into a secure development cycle or SOC means entrusting Anthropic with data regarding vulnerabilities within one’s infrastructure. As a U.S. entity subject to the Cloud Act, this configuration necessitates a specific risk analysis prior to any deployment for vital European operators and entities under NIS2 obligations. The regulatory exposure does not only concern data processed in production but also the sensitivity of the vulnerability information itself. The defensive power of Glasswing and the legal constraints of sovereignty are two sides of the same coin, and European CISOs must navigate both simultaneously.