Visa open-sources VVAH, its Glasswing agentic security harness

Visa just published something more interesting than another scanner. The Visa Vulnerability Agentic Harness, or VVAH, is an Apache 2.0 pipeline the payments network built on learnings from Anthropic's Project Glasswing, and it is worth reading past the badges.

The framing is the part I want to flag. Most AI-security tooling still pitches on discovery, more findings per scan, more novel bug classes. VVAH's README argues the real bottleneck is Mean Time to Adapt, defined as the elapsed time from AI-discovered exploitability to a validated fix in production. The pipeline runs eleven stages across four phases: discovery and modeling, deep dive and adversarial verification, synthesis and structured reporting, then remediation and validation. Different agent roles handle each half, with discovery and offensive agents on the finding side and developer and validation agents on the fix side. Deterministic multi-agent voting is used to knock down false positives before a human is asked to look.

Why this matters if you run application security at any regulated shop: the honest problem with LLM-driven vuln research has never been the finding, it is the triage queue that swamps whoever has to read the output. A card network open-sourcing its own opinion on the right ratio of discovery to remediation is a useful anchor, and the SARIF 2.1.0 output means it slots into the tooling teams already run rather than asking for another portal.

The caveats are also in the README and they matter. There are no published precision or recall numbers, token consumption is called out as high without a figure attached, and the remediation and validation stages currently require an Anthropic backend even though the earlier detection stages are vendor-neutral. What the reporting does not give you is how much VVAH has been run against Visa's own production code, or how many of the fixes it proposed were actually shipped.

Still, the direction is the interesting part. If regulated buyers start measuring their security AI on Mean Time to Adapt rather than finding count, the market pressure on scanners changes shape, and vendors who cannot close the loop to a validated fix start looking incomplete.