Visa uses Anthropic's AI to hunt bugs in its payment network and releases tool

What happened?

Visa used Anthropic's Mythos system, based on its Claude model, to conduct an automated vulnerability hunt on its own payment network, which processes transactions in over 200 countries, 160 currencies, and connects nearly 5 billion credentials to more than 175 million merchants. According to VentureBeat, the test was part of Anthropic's Glasswing project, where several organizations tested Mythos on critical systems. In the first month, Glasswing identified over 10,000 high or critical severity vulnerabilities across the sector.

Rajat Taneja, Visa's president of technology, presented the results at VB Transform 2026 and detailed that Mythos demonstrated system-level contextual analysis, finding deep vulnerabilities in the technology stack and chaining them into functional exploit paths. Visa attributes its zero-trust controls and network segmentation to breaking the chain before a real attacker could exploit them. Taneja, who has led Visa's technology strategy since 2019 after 15 years at Microsoft and a stint as CTO of Electronic Arts, co-authored the June 10 blog and technical white paper with CISO Subra Kumaraswamy.

As a result, Visa published the Visa Vulnerability Agentic Harness on GitHub, a test harness that governed the entire search, as an open source reference for any security team to inspect, adapt, and extend. It also published a technical white paper with the architecture, lessons learned, and 12 non-negotiable architectural practices for critical infrastructure.

Why is it important?

This case marks a milestone in the use of agentic artificial intelligence for offensive and defensive cybersecurity. Traditionally, SAST (Static Application Security Testing) tools detect known patterns, but Mythos goes further by chaining seemingly minor weaknesses into full exploits, something only achieved in late stages of manual penetration testing. As Taneja noted: “In a world of agentic attacks, defense must also be agentic.” This approach contrasts with earlier methods, such as rule-based vulnerability scanners, which often generate false positives and fail to consider system context. Mythos, instead, understands the complete architecture and can simulate real attack paths.

Additionally, Visa abandoned traditional remediation metrics (like number of closed vulnerabilities) for its own measure that prioritizes the real risk of exploit chains. This shift could influence how other companies measure the effectiveness of their security teams, aligning with standards like CVSS v4 that also weigh exploit context. Visa's decision to release the harness as open source is also significant: it aims for the community to adopt similar practices, accelerating innovation in automated cybersecurity and setting a transparency standard.

Consequences and lessons

• Agentic AI maturity: Mythos demonstrates that advanced language models can act as autonomous security agents, reducing the time between vulnerability detection and exploitation. This could displace manual penetration testing in critical environments, though human oversight is still needed to avoid catastrophic errors.
• Open source as standard: By releasing the harness, Visa seeks for the community to adopt similar practices, accelerating innovation in automated cybersecurity. This move echoes the release of tools like Metasploit, which democratized penetration testing, but now with an agentic AI focus.
• Metric shift: Visa's new metric, based on exploit chains rather than isolated vulnerabilities, could become an industry reference. Companies like Google and Microsoft have already experimented with similar metrics in their bug bounty programs, but Visa applies it at the critical infrastructure level.
• Implications for critical infrastructure: Visa's white paper offers 12 non-negotiable architectural practices that other critical infrastructure companies should consider. These include network segmentation, continuous monitoring, and automated response, principles already common in sectors like banking or energy, but now reinforced by AI's ability to find weaknesses in complex chains.

An additional detail: Taneja emphasized that trust in Visa is built on pessimism and paranoia, assuming failures and designing around them. The network has been hardened for years with zero-trust architecture, layered defenses, and highly automated security operations. This allowed the vulnerabilities found by Mythos to be unexploitable in production, as controls broke the attack chains.

What should readers know?

For security professionals, this case underscores the need to adopt agentic AI tools to complement traditional analyses. Tools like Mythos can reduce detection time for critical vulnerabilities from weeks to days, but require a solid defense infrastructure (zero trust, segmentation) so findings don't become real risks. For executives, it shows that investment in defense in depth remains essential, even when deploying advanced technologies. The open source harness and white paper allow any organization to assess its own security posture without incurring licensing costs.

For the general public, it's a sign that payment companies are using AI to protect their data, though it also raises questions about reliance on third-party models (Anthropic) and the possibility that these same systems could be used by attackers. However, Visa mitigated this risk by publishing the harness as a reference, allowing the community to contribute to improving collective security. The white paper details how to integrate these systems without compromising network stability, including the need for sandboxing and scope limitations.

In summary, the Visa and Mythos case represents a significant advance in agentic cybersecurity, with implications ranging from risk measurement to open source transparency. The industry will be watching whether other companies follow Visa's lead and adopt exploit-chain-based metrics, and whether agentic AI becomes a standard for critical infrastructure security.