Visa directed the capabilities of Anthropic’s Claude Mythos towards its transaction infrastructure, which handles billions of transactions daily across more than 200 countries and territories. This system processes transactions in approximately 160 different currencies, linking almost 5 billion payment credentials with over 175 million merchant locations.
This model was able to detect minor vulnerabilities within the system and connect them into exploit chains, which typically only emerge during late stages of penetration testing. Rajat Taneja, Visa’s president of technology, shared this information with the audience at VB Transform 2026, explaining why Visa decided to open-source the tool governing this process and why they moved away from traditional remediation metrics in favor of a new measurement developed by their team.
Since 2019, Taneja has been responsible for technology strategy, product engineering, and global infrastructure at Visa. He joined Visa in 2013 after working at Electronic Arts as CTO and spending 15 years at Microsoft. Alongside Visa’s chief information security officer, Subra Kumaraswamy, Taneja co-authored a blog post on June 10 announcing the release of the Visa Vulnerability Agentic Harness on GitHub, designed as a reference that security teams can inspect, adapt, and extend. Visa also released a technical white paper that outlines the architecture, lessons learned, and 12 crucial architectural practices for critical infrastructure.
Trust Built on Pessimism and Paranoia
Taneja highlighted the rationale for Visa’s rigorous defense measures, emphasizing that trust on a global payment scale is built by assuming failure is inevitable and designing systems to withstand it. This approach, described as pessimism and paranoia, has led to the network’s reinforcement over years through zero-trust architecture, multiple defense layers, and automated security operations essential for global payment reliability.
When Anthropic invited Visa to test Mythos under Project Glasswing, Visa agreed. During the first month of testing, according to Anthropic, participants identified over 10,000 high- or critical-severity vulnerabilities in software across various industries. Anthropic noted a bottleneck in verification, disclosure, and patching speed. Visa’s involvement aimed to test their system’s robustness at AI speed and see where advanced models could further enhance their defenses.
Insights from Mythos at Visa
Within Visa’s system, Mythos provided comprehensive, context-aware analysis, uncovering deeply embedded vulnerabilities and highlighting issues that become more severe when linked. The findings were sufficiently clear for engineering teams to address without unnecessary complexity. Some findings were critically severe, but Visa’s zero-trust controls, network segmentation, and layered safeguards prevented exploitation.
While this validation was important, Taneja pointed out that the subsequent realization was even more significant. “In a world of agentic attacks, defense also has to be agentic,” he stated. Despite extensive investment in defense-in-depth, the model highlighted assumptions that needed reevaluation. Visa’s white paper explains that traditional SAST tools remain useful for identifying known vulnerability patterns, but can’t anticipate adversaries who exploit logic, data flow, and exploit chains.
A Harness, Not a Scanner
Visa’s solution wasn’t just another scanner. The Visa Vulnerability Agentic Harness, now in its fifth version, directs advanced AI models through structured security tasks with deterministic controls, policy gates, and human oversight. Taneja outlined its design, which spans four phases and eleven stages, from code ingestion to threat modeling, deep verification, exploit chain synthesis, and ultimately remediation and validation.
The project’s documentation highlights three design choices that enhance finding quality: threat modeling precedes analysis to target the attack surface, multi-agent deterministic voting requires consensus across reasoning chains before advancing a finding, and structured triage artifacts streamline the process from discovery to actionable results. The pipeline defaults to a high operational pace. A standard scan executes all eleven stages, updating source files in the target repository in fix mode, applying patches unless intervention occurs during detection.
The harness is designed for multiple models. An LLM abstraction layer allows Visa to swap or combine providers without altering the control plane. The open-source version supports Anthropic Claude, OpenAI-compatible models, or a combination. However, remediation and validation stages require Anthropic models for full functionality, as they provide the necessary file-editing tools, while OpenAI-compatible models are limited to reporting. VentureBeat’s Q2 2026 Pulse research, presented earlier, emphasized the importance of provider flexibility, with 82% of surveyed enterprises relying on provider-native controls as their main security layer, and 59% planning to change agent security tooling within a year.
Mean Time to Adapt Replaces Legacy Metrics
Taneja argued that finding vulnerabilities is no longer the main challenge. The true test lies in verifying an issue’s exploitability, resolving it, and confirming the closure of the attack path beyond just applying a patch. Visa calls this Mean Time to Adapt (MTTA), tracked in the white paper along three dimensions: inventory freshness, exploitable paths per release, and validation cycle time. MTTA emphasizes the importance of these outcomes, as opposed to legacy measures like mean time to detect and CVE closure counts, which can mask underlying exposure. Visa’s SSDLC policy assumes every exploitable path will be utilized in production and mandates remediation before code promotion.
AI Accelerates Supply Chain Risk
Moving beyond Visa’s internal systems, Taneja addressed the supply chain risks. The white paper warns that a well-protected enterprise remains vulnerable through weak vendors and open-source components. Therefore, Visa is enforcing AI-specific security measures as a crucial part of supplier evaluations, requiring ongoing vulnerability validation, updated software bills of materials, and MTTA baselines throughout its technology stack.
Visa has also joined Project Lightwell, a $5 billion initiative by IBM and Red Hat to strengthen widely used open-source components through AI-driven validation and coordinated patching, alongside financial institutions like Bank of America, JPMorganChase, Goldman Sachs, and Mastercard. This initiative extends the MTTA logic beyond individual companies.
When Agents Begin Transactions
Visa’s next challenge is securing transactions conducted by AI agents. Taneja envisioned a future where AI agents act on behalf of consumers and businesses, stating that Visa is developing the necessary trust framework, identity layer, and agent readiness assessments for safe transactions. The Visa Payment Threats Lab, a simulation environment, analyzes real fraud scenarios against Visa’s actual authorization rules to identify AI-enabled vulnerabilities and recommend targeted defenses.
The identity challenge is concrete. VentureBeat’s Pulse research found 69% of enterprises use credential sharing in their agent deployments, with a 63.5% incident rate for shared credentials versus 40.9% for individually scoped identities. Visa’s white paper addresses this by listing “AI agents are identities” among its 12 essential practices, requiring scoped permissions, least privilege enforcement, full audit trails, and IAM governance inclusion for every agent interacting with APIs, reading data, or modifying systems.
Three Defensive Priorities
Visa’s defensive strategy focuses on three priorities, according to Taneja. These include shifting security efforts left to eliminate exploitable flaws before production, replacing high-risk, unsupported components to prevent exposure, and refactoring defenses to operate autonomously with human oversight in response to increasing threat volumes and evolving attack models.
Implementing these strategies doesn’t require a large budget. The harness is available on GitHub, with 595 stars and 97 forks as of July 20, and MTTA can be tracked via a dashboard without a lengthy procurement process. The white paper’s 12 essential practices align with existing security architecture reviews. Visa’s paper concludes with a sense of urgency, stating that the chance to outpace machine-speed attackers is still present, but it won’t last indefinitely.
