Visa Vulnerability Agentic Harness opens after Mythos security test

Visa has open-sourced the Visa Vulnerability Agentic Harness after using Anthropic’s Claude Mythos to probe parts of the infrastructure that supports its global payments network. The move matters because Visa is putting a governed AI security workflow into public view after testing it against systems tied to billions of daily transactions.

Rajat Taneja, Visa’s president of technology, described the work at VB Transform 2026. Visa also published a June 10 blog post, co-authored by Taneja and chief information security officer Subra Kumaraswamy, along with a technical white paper and the project’s GitHub repository.

Visa says its network spans more than 200 countries and territories, handles roughly 160 currencies, and connects nearly 5 billion payment credentials with more than 175 million merchant locations. That scale explains why the company frames security around pessimistic assumptions: systems will fail, attackers will chain weaknesses, and controls must limit damage before exploitation reaches customers.

What is the Visa Vulnerability Agentic Harness?

The Visa Vulnerability Agentic Harness is an open-source reference implementation for directing frontier AI models through vulnerability discovery, verification, exploit-chain analysis, remediation and fix validation. Visa describes it as a controlled pipeline rather than a standalone scanner, with policy gates, deterministic controls and human oversight built into the process.

According to Visa’s documentation, the harness runs across four phases and 11 stages, starting with code ingestion and threat modeling before moving into deeper validation and remediation workflows. The project is now in its fifth generation, Visa said.

The harness supports Anthropic Claude and OpenAI-compatible models through an abstraction layer. Visa’s documentation also notes a limitation: source-file editing for automated fixes currently depends on Anthropic backends, so OpenAI-compatible models in those stages are limited to report-only output.

What did Mythos find at Visa?

Visa said Mythos identified weaknesses buried in its stack and showed how smaller issues could combine into more serious exploit paths. Some findings were rated critical, according to the company, though Visa did not disclose the number of findings, affected systems or remediation timeline.

Taneja said Visa’s existing zero-trust controls, network segmentation and layered safeguards interrupted the chains before an outside attacker could have used them. That is Visa’s claim; the company did not publish exploit details, which is typical for critical infrastructure security work.

The test was part of Anthropic’s Project Glasswing, which invited organizations tied to critical software to test Mythos. Anthropic has said Glasswing participants found more than 10,000 high- or critical-severity vulnerabilities in the first month across critical systems software industry-wide. Anthropic also said the bottleneck shifts after discovery, into verification, disclosure and patching.

Why Visa changed the metric

Visa is promoting a metric it calls Mean Time to Adapt, or MTTA, instead of relying mainly on mean time to detect or raw CVE closure counts. The company defines MTTA around three measures: how current an organization’s inventory is, how many exploitable paths remain after each release, and how long it takes to prove a fix closes the attack route.

The argument is straightforward for security teams: closing tickets can leave exploitable chains intact if validation is shallow. Visa’s white paper cites CISA Known Exploited Vulnerabilities data to argue that fewer than 1% of CVEs are actively exploited, making exploitability a more useful prioritization signal than volume alone.

Visa says its secure software development lifecycle now assumes every exploitable path will be tried in production and requires remediation before code promotion. That is a high bar, and Visa did not say how broadly the policy has been implemented across all internal systems and suppliers.

Why open source the harness?

Visa said it released the harness so other security teams can inspect, adapt and extend the workflow. The GitHub project had 595 stars and 97 forks as of July 20, according to the published account.

The supplier angle is central to Visa’s strategy. The company says it is making AI-specific security posture part of vendor due diligence, including continuous vulnerability validation, current software bills of materials and MTTA baselines. Visa has also joined Project Lightwell, IBM and Red Hat’s $5 billion effort to harden widely used open-source components with AI-driven validation and coordinated patching, alongside Bank of America, JPMorganChase, Goldman Sachs and Mastercard.

Taneja also tied the work to agentic commerce, where AI agents transact for consumers or enterprises. Visa says it is building trust, identity and agent-readiness systems for that scenario, and its white paper treats AI agents as identities that need scoped permissions, audit trails and IAM governance.

This story draws on original reporting from VentureBeat.