Visa tested Anthropic’s Claude Mythos against infrastructure supporting its global payment network, uncovering vulnerability chains that conventional security tools might have identified only during later penetration testing.

The network operates across more than 200 countries and territories, handles roughly 160 currencies, and connects nearly 5 billion payment credentials with more than 175 million merchant locations.

During the test, Claude connected smaller weaknesses buried across different parts of Visa’s technology stack into potential attack paths. Some findings received critical severity ratings.

Visa said its existing zero trust architecture, network segmentation, and layered security controls prevented the identified weaknesses from becoming complete attack paths that an external attacker could exploit.

The test was conducted through Project Glasswing, an Anthropic initiative that invited organizations responsible for critical software and infrastructure to evaluate advanced AI models against their systems.

Project participants identified more than 10,000 high or critical severity vulnerabilities across critical software during the first month of testing, according to Anthropic.

Visa president of technology Rajat Taneja said the results showed that security teams will need automated defenses capable of operating at the same speed as AI driven attacks.

“In a world of agentic attacks, defense also has to be agentic,” Taneja said during VB Transform 2026.

Following the test, Visa developed the Visa Vulnerability Agentic Harness, a governed system that directs advanced AI models through structured security investigations.

The harness operates across four phases and eleven stages, covering code ingestion, threat modeling, vulnerability analysis, verification, exploit chain construction, remediation, and fix validation.

Rather than giving an AI model unrestricted control, the system combines model reasoning with deterministic safeguards, policy controls, and human oversight.

Threat modeling is completed before deeper analysis begins, helping the system focus on relevant attack surfaces. Multiple independent reasoning processes must also reach similar conclusions before a vulnerability advances through the pipeline.

Visa said this approach reduces false positives and produces findings that engineering teams can use without extensive manual review.

The harness can generate structured security reports, proposed code changes, and evidence showing whether a vulnerability has actually been resolved.

Visa released the system on GitHub as an open source reference implementation that security teams can inspect, modify, and extend.

The architecture supports Anthropic models, OpenAI compatible models, and combinations of different providers. However, full remediation currently requires file editing tools available through Anthropic models. Other compatible models are limited to analysis and reporting in those stages.

Visa also introduced a new measurement called Mean Time to Adapt.

The company argues that traditional metrics, including detection speed and the number of closed vulnerabilities, do not show whether an organization has eliminated the underlying attack path.

Mean Time to Adapt instead measures how quickly a team can confirm that a weakness is exploitable, apply a fix, and prove that the complete attack chain has been closed.

The framework tracks the freshness of an organization’s software inventory, the number of exploitable paths remaining after each release, and the time required to validate a fix in production.

Visa said its software development policy now assumes every exploitable path could eventually be used and requires those paths to be addressed before code reaches production.

The company is also applying similar standards to suppliers and open source components as AI models make it easier to identify and combine weaknesses across connected systems.