Quick Facts
- Visa’s Vulnerability Agentic Harness (VVAH) now discovers, patches, and validates code fixes end-to-end with no required human review before changes are applied.
- The tool runs a four-phase, eleven-stage pipeline; in default configuration it edits source files before any human reads the finding or the patch.
- VVAH is built on learnings from Anthropic’s Project Glasswing, where the Mythos AI model identified 6,202 high- or critical-severity vulnerabilities across open-source projects.
Visa on Aug. 27 announced an expanded version of its Visa Vulnerability Agentic Harness, an open-source security tool that writes, applies, and validates code patches with no human gate between discovery and fix. The update adds remediation and validation to the tool’s original vulnerability discovery capability.
Rajat Taneja, Visa’s President of Technology, said the threat timeline left no alternative. “AI is compressing the time between vulnerability discovery and exploitation, which means defenders need a faster, more reliable path to action,” he told VentureBeat.
VVAH runs as a four-phase, eleven-stage pipeline. The first three phases map the attack surface, confirm exploitability through adversarial verification, and produce structured findings. The final two stages propose a fix and run an adversarial validation panel against it before the patch is flagged for adoption. Visa says some resolutions have shrunk from weeks to hours.
The tool traces back to Visa’s participation in Anthropic’s Project Glasswing, a frontier AI cybersecurity initiative. Inside that program, Anthropic’s Mythos model identified a 27-year-old vulnerability in OpenBSD and a 16-year-old flaw in FFmpeg, both in code that had already passed extensive automated testing. Across open-source projects, Mythos found an estimated 6,202 high- or critical-severity vulnerabilities out of 23,019 total findings.
Carl Rutstein, Visa’s Global Head of Consulting and Analytics, framed the product around speed. “Finding vulnerabilities is no longer the hardest part,” he said. “Speed to remediation is the new battleground.”
The default behavior of VVAH creates a governance question that engineering and security leaders will need to address directly. In its standard configuration, the tool edits source files on a developer machine or build agent before any human has reviewed the finding. Visa notes the tool does not touch live production systems and that no change is merged, signed, or deployed without a human action. But the full find-fix-validate sequence runs without a human checkpoint in the middle.
Full file-editing capability requires an Anthropic backend. Teams running VVAH against an OpenAI-compatible endpoint receive report-only output. That means the choice of AI model also determines the tool’s write permissions, a distinction Visa does not surface prominently in its configuration documentation.
The autonomy creates tension with Visa’s own stated security principles. The company’s June white paper lists “AI agents are identities” as one of 12 non-negotiable practices, requiring scoped permissions, least-privilege access, audit trails, and IAM governance for every agent that modifies a system. Whether VVAH’s default behavior meets those standards in practice is a question each deploying organization will have to answer for itself.
Taneja pointed to the Hugging Face incident and AI models escaping sandboxes as evidence of what is coming. “We have seen the trailer of this movie,” he said, adding that every company should rethink its architecture now.
Visa positioned VVAH as a tool that handles repetitive security work so that human engineers can focus on decisions requiring judgment. “VVAH does the repetitive parts,” Visa wrote in its documentation. “Before a fix gets to a human, an adversarial validation panel at stage eleven tries to break it.” The company said its goal is to help security teams work at AI speed, not to replace them.
For founders and executives running software companies, VVAH signals where enterprise security is heading: autonomous agents that act first and report to humans second. The critical decision is not whether to adopt such tools, but how to set the permission boundaries before the pipeline runs.
Read more: Visa ships a security AI that patches production code before any human reviews it