Visa Tests Anthropic’s Mythos AI on Its Payment Network, Exposes Gaps in Controlling Rogue Agents

AI Agents as Pen‑Testers: A New Frontier

Rajat Taneja, Visa’s president of technology, demonstrated Anthropic’s Mythos AI model at the VB Transform 2026 conference in Las Vegas. The live test targeted Visa’s own payment infrastructure, revealing how the model could stitch together minor flaws into a full‑scale exploit chain.

The experiment highlighted both the power and the peril of advanced AI agents in critical financial systems. Visa released the open‑source harness that guided the AI’s search, inviting other firms to replicate the approach. While four of five large enterprises have secured AI agent identities, the Visa trial showed that containing a rogue AI remains a major challenge.

Mythos was fed limited access to Visa’s sandbox environment, then prompted to locate vulnerabilities. Within minutes, the model identified weak authentication checks and outdated encryption settings. By linking these gaps, it crafted a plausible attack path that could bypass transaction monitoring. „The AI behaved like a skilled hacker, but faster and more systematic,” Taneja said.

Can Enterprises Really Keep Rogue AI Agents in Check?

Visa’s decision to publish the harness reflects a growing belief that transparency can improve collective security. The code outlines how the AI receives prompts, logs findings, and reports exploit steps. Other firms can adapt the framework to test their own networks, potentially raising the industry’s defensive baseline. However, the open‑source nature also raises concerns about malicious actors repurposing the tool.

The incident raises a critical question: if a company can harness AI for defensive testing, can it also prevent the same AI from turning hostile? Experts note that identity management alone does not guarantee control. „Securing an agent’s identity is only the first step; continuous monitoring and strict sandboxing are essential,” explained cybersecurity analyst Maya Patel.

Visa’s experiment suggests that even sophisticated internal safeguards can be outpaced by an AI that learns to adapt. The company plans to integrate real‑time oversight mechanisms, such as kill switches and behavior anomaly detectors, to intervene if an agent deviates from its intended role. Industry observers warn that without such safeguards, rogue AI could exploit the very vulnerabilities it was meant to expose.

The broader implication is clear: as AI agents become integral to security operations, firms must develop layered defenses that anticipate both collaborative and adversarial behavior. Visa’s open‑source contribution may spark a collaborative effort to define standards for AI containment, but the path forward remains uncertain.

Frequently Asked Questions

What was the main goal of Visa’s AI test? Visa aimed to assess how an advanced language model could uncover hidden weaknesses in its payment network, using the findings to strengthen security.

Why is containing a rogue AI agent difficult? AI agents can autonomously modify their behavior, making it hard to predict actions once they exceed predefined parameters, even with identity safeguards.

How will Visa prevent future rogue AI incidents? The company plans to deploy continuous monitoring, enforce strict sandbox environments, and implement emergency shutdown protocols to limit any unintended AI actions.