How Visa Uses Frontier AI to Protect Global Payments 🛡️

Processing billions of daily transactions across more than 200 countries and 160 currencies requires an architecture built on absolute resilience. Connecting nearly 5 billion payment credentials to over 175 million merchant locations, Visa operates one of the world's most critical financial infrastructures. At this scale, cyber security is engineered through what Visa's technology leadership calls "pessimism and paranoia"—assuming system failures will occur and designing multi-layered safeguards before threats ever materialise.

To test decades of security hardening against next-generation threats, Visa recently put frontier artificial intelligence directly against its core network infrastructure. Participating in Anthropic’s Project Glasswing initiative, Visa deployed Anthropic's advanced model, Claude Mythos, to audit its codebase. The results not only transformed Visa's internal defensive strategy, but also led to the open-sourcing of a brand-new security tool for the broader developer community.

When AI Uncovers Deep Exploit Chains

Traditional security scanners—such as Static Application Security Testing (SAST) tools—rely heavily on pattern matching to flag known vulnerability signatures. While effective for initial code reviews, pattern matching often struggles to detect logical flaws that exist between isolated components.

When subjected to Claude Mythos, Visa’s network demonstrated the power of context-aware, system-wide analysis. The AI model identified subtle, low-severity weaknesses buried deep within the stack and demonstrated how an attacker could stitch them together into working exploit chains.

Because Visa operates under a strict zero-trust model with aggressive network segmentation, these theoretical chains were broken internally before any threat actor could exploit them. However, the experiment revealed a vital strategic truth: in an era where cyber attackers utilise autonomous AI agents, cybersecurity defences must become agentic as well.

Introducing the Visa Vulnerability Agentic Harness

Rather than building another static scanning tool, Visa engineered a governed pipeline designed to direct frontier AI models through structured defensive tasks while maintaining strict human oversight and deterministic controls. Visa has now open-sourced this technology, releasing the Visa Vulnerability Agentic Harness on GitHub as a reference architecture for security teams worldwide.

The harness operates across four structured phases and eleven distinct stages, spanning code ingestion, threat modelling, deep verification, exploit synthesis, and candidate patch validation:

• Threat Modelling First: Analysis begins by mapping out actual attack surfaces rather than scanning entire repositories blindly.
• Multi-Agent Voting: To eliminate noise and false positives, independent AI reasoning chains must reach convergence before a finding is flagged.
• Automated Remediation: Operating in fix mode, the pipeline can edit source files directly in the target repository, generating candidate patches that developers can rapidly review and ship.

Designed with model-agnostic flexibility, the framework features a language model abstraction layer. Security teams can pair the harness with Anthropic Claude, OpenAI-compatible models, or hybrid configurations. Full automated remediation currently relies on file-editing tools supported by Anthropic backends, while other models provide comprehensive report-only outputs.

Shifting from Legacy Metrics to Mean Time to Adapt (MTTA)

A core takeaway from Visa's research is that finding vulnerabilities is no longer the primary hurdle in security operations; confirming exploitability and applying verified fixes quickly is what truly counts. Industry data shows that under 1% of published Common Vulnerabilities and Exposures (CVEs) are ever actively exploited in production. Consequently, chasing raw CVE closure counts can create a false sense of security while leaving critical exploit chains open.

Visa is replacing legacy metrics like Mean Time to Detect (MTTD) with a new operational benchmark: Mean Time to Adapt (MTTA). MTTA evaluates defence readiness across three core dimensions:

1. Inventory Freshness: How complete and real-time the organisation’s tracking of code, configuration, and active deployments is.
2. Exploitable Paths per Release: Tracking the net reduction of end-to-end attack chains after each deployment, rather than simple flaw counts.
3. Validation Cycle Time: The speed at which a security team can generate repeatable, evidence-backed proof that a patch permanently closes an attack path.

Under Visa’s revised Secure Software Development Life-cycle (SSDLC) policy, every exploitable path identified must be remediated before code can be promoted to production.

Managing Third-Party Risk and the Future of Agentic Commerce

The shift toward machine-speed security extends far beyond an enterprise's immediate perimeter. Weak points within vendor supply chains and open-source dependencies represent significant exposure risks. To combat this, Visa requires suppliers to meet rigorous AI security baselines, maintain living Software Bills of Materials (SBOMs), and demonstrate clear MTTA standards. Additionally, Visa has joined Project Lightwell—a $5 billion open-source hardening initiative led by IBM and Red Hat alongside major global financial institutions.

Looking ahead, Visa is actively preparing its infrastructure for the rise of autonomous agentic commerce, where AI agents will make purchases on behalf of consumers and businesses. This transition introduces complex identity management challenges. Research shows that enterprise credential sharing among AI agents dramatically increases security incidents.

To address this, Visa enforces the principle that AI agents are distinct identities. Every agent interacting with APIs or databases must operate under scoped permissions, least-privilege access rules, and complete Identity and Access Management (IAM) governance. Through the Visa Payment Threats Lab, real-world fraud scenarios are continually simulated to stress-test authorisation logic against emerging AI threats.

Three Priorities for Modern Security Teams

Visa outlines three strategic imperatives for defenders aiming to stay ahead of machine-speed adversaries:

• Shift Left Aggressively: Design out exploitable flaws during the early development phase before code ever reaches production environments.
• Eliminate High-Risk Dependencies: Proactively replace unmaintained or vulnerable open-source components across the technology stack.
• Deploy Governed Autonomous Defences: Refactor security operations so that automated detection, validation, and patch generation keep pace with rapid threat escalation.

By releasing its agentic harness and detailing its architectural practices, Visa has provided a blueprint that organisations of any size can adapt to harden their own digital infrastructure.

Further Reading: 

For full technical details, white paper documentation, and further insights on this development, read the original report on VentureBeat:

👉 Visa used Mythos to hunt for bugs in its own payment network, then open-sourced the harness that made it possible

Disclaimer: This article is provided for informational purposes only, mistakes may be made, and it's not offered or intended to be used as legal, tax, investment, financial, or any other advice.