Analysis
Visa released the Visa Vulnerability Agentic Harness (VVAH) in full on Aug. 27, an open-source security system that runs a discover, verify, remediate, validate, iterate loop across production code. VentureBeat detailed the architecture: eleven stages, with Stage 10 writing candidate fixes and Stage 11 performing adversarial validation before anything reaches a human reviewer. It builds abstract syntax tree call graphs to map code paths and pulls in CMDB data, threat models and business risk context to prioritize.
The framing from Visa's president of technology Rajat Taneja is the part worth sitting with: "The bottleneck has moved. AI is finding vulnerabilities faster than humans can." And: "It's not the finding. It's the fixing that matters." That is a genuine inversion of the last decade of security tooling, which optimized detection and generated queues no team could work through. Carl Rutstein, global head of Visa Consulting and Analytics, put it as "speed to remediation is the new battleground."
The multi-model design is the interesting engineering choice. Rather than standardizing on one model, VVAH selects per stage based on precision and recall characteristics -- Claude Mythos where high recall matters, Claude Opus where precision matters, GPT and open-weight models configurable through OpenAI-compatible backends. That is what a mature production AI system looks like, and it is the opposite of the single-vendor architecture most enterprises started with.
“Adoption is real if early: 595 GitHub stars on July 20, more than 2,300 stars and 300-plus forks by Aug.”
Adoption is real if early: 595 GitHub stars on July 20, more than 2,300 stars and 300-plus forks by Aug. 25, with a clone-to-visitor ratio around 9% -- meaning people are running it, not just bookmarking it. Visa reports mean time to adapt on some resolutions shrinking from weeks to hours.
The counter-argument came from Steve Wilson, chief AI and product officer at Exabeam: "The first thing I'd do is put an authorization gate outside the model." He is right. An agent with write access to production code is an extraordinarily attractive target, and the Tenet Security GhostJacking demonstration at DEF CON 34 eighteen days before this announcement showed how agent pipelines get hijacked. Pulse covered a parallel case this week -- Claude Code tricked by a prompt injection hidden in a webpage it was asked to summarize.
A payments network open-sourcing its internal security tooling is also a competitive act. It sets a reference architecture that Mastercard, the card networks and every bank security org now has to respond to.
