Security teams have long operated in a reactive cycle, chasing known signatures and patching holes after they are flagged by static analysis tools. For decades, the industry has relied on the hope that a comprehensive set of rules could anticipate every possible attacker move. But as the attack surface expands across global cloud infrastructures, the gap between finding a bug and understanding how that bug can be weaponized has become a critical liability. This week, the conversation shifted from simple detection to autonomous reasoning as one of the world's largest payment processors revealed how it used frontier AI to stress-test its own defenses.

The Scale of Project Glasswing and Claude Mythos

Visa recently integrated Claude Mythos, a security-specialized model from Anthropic, into its global payment network infrastructure. This effort was part of Project Glasswing, a collaborative software security testing initiative involving companies that maintain the core systems underpinning modern industry. The goal was to determine exactly how far a high-performance AI model could push existing defense mechanisms. The results were immediate and staggering: within the first month of testing, the system identified more than 10,000 vulnerabilities categorized as high or critical.

This was not a localized test. Claude Mythos was deployed across a network spanning more than 200 countries and territories, analyzing the systems that process billions of transactions daily. By leveraging the computational speed of AI, Visa was able to compress decades of security hardening and verification into a fraction of the time. To manage this process, Visa developed the Visa Vulnerability Agentic Harness, a framework designed to govern how AI models explore and identify weaknesses. Recognizing the broader need for such a system, Visa has released the reference implementation of the harness as an open-source project on GitHub, allowing other security organizations to build their own AI-driven vulnerability discovery pipelines.

From Pattern Matching to Exploit Chain Reasoning

To understand why this matters, one must look at the failure of traditional Static Application Security Testing (SAST). Conventional SAST tools operate on a first-pass basis, scanning source code for known vulnerability patterns. They are essentially sophisticated search engines for signatures. While effective at finding isolated errors, they are blind to the logic of a human attacker. An attacker does not look for a single bug; they look for a chain of minor weaknesses that, when linked together, create a path to a critical system. Traditional tools cannot reason through these data flows or infer the logical connections between disparate flaws.

Claude Mythos changes this dynamic by shifting the paradigm from pattern matching to logical inference. Instead of flagging a line of code as suspicious, the model reasons through the stack, identifying minor weaknesses and synthesizing them into a functional exploit chain. This ability to simulate the attacker's logic allows the AI to uncover vulnerabilities that were previously invisible to signature-based tools.

To ensure this autonomy does not lead to chaos, the Visa Vulnerability Agentic Harness implements a rigorous governance pipeline consisting of four phases and 11 distinct stages. The process begins with code collection and threat modeling, moves into deep verification, proceeds to exploit chain synthesis, and concludes with remediation and validation. Every stage is guarded by deterministic controls, policy gates, and human oversight. This structure ensures that while the AI provides the reasoning power, the organization maintains absolute control over the execution.

Technical flexibility is baked into the architecture via an LLM abstraction layer. This allows security teams to swap or combine model providers without modifying the underlying control plane. The open-source version supports Anthropic Claude, OpenAI-compatible models, or a hybrid mix of both. However, a critical technical distinction exists regarding the remediation phase. The ability to actually edit files and validate fixes requires specific tool-use capabilities exposed only in the Anthropic backend. While OpenAI-compatible models can be used for the discovery and reporting phases, they are limited to outputting reports because they lack access to the necessary file-editing tools required for full autonomous remediation.

This shift in capability necessitates a shift in how success is measured. The industry has traditionally focused on the number of patches applied, but Visa argues that the real metric is MTTA, or Mean Time to Adapt. MTTA measures whether the actual attack path has been blocked, rather than simply whether a piece of code was changed. By focusing on the path rather than the patch, organizations can finally measure the actual effectiveness of their security posture in the face of agentic threats.

The release of the Agentic Harness signals a move toward a future where security is an autonomous, continuous loop of attack and defense.