Visa taps Anthropic AI to uncover 10,000 security flaws in global payment network

Visa’s payment system, which operates in more than 200 countries and processes transactions in roughly 160 currencies, used Anthropic’s Claude Mythos model in a project called Glasswing that began in early April 2026. Within the first month the AI identified over 10,000 high‑ or critical‑severity vulnerabilities across Visa’s software stack, linking minor issues into full exploit chains that traditional scans miss.

Visa’s president of technology, Rajat Taneja, presented the findings at VB Transform 2026 and introduced a new metric, “Mean Time to Adapt,” to replace conventional remediation timelines. The company’s zero‑trust architecture prevented any of the discovered flaws from being exploited. On 10 June 2026 Visa released the Visa Vulnerability Agentic Harness (VVAH) as an open‑source project on GitHub, enabling other organisations to use the same AI‑driven approach. Anthropic supported the effort with up to $100 million in credits and an additional $4 million for the open‑source work.