Visa open-sources VVAH: AI finds the bug and drafts the fix

Finding bugs stopped being the hard part. Visa's Vulnerability Agentic Harness (VVAH) now runs past discovery into proposing patches and grading them — and the whole thing is on GitHub under Apache 2.0. Before you read the headline version of this, read the repo. They disagree.

What actually happened

VVAH came out of Project Glasswing, Visa's defensive security program built on Anthropic's Mythos model. In Visa's own writeup, Glasswing participants flagged more than 10,000 high- or critical-severity vulnerabilities in widely used software in the first month. Visa's conclusion was blunt: security is no longer limited by how fast you find things, but by how fast you verify, disclose, and patch them.

The August 27 release extends the pipeline from discover, verify, report to remediate, validate, iterate. It is eleven stages across four phases. S1–S3 map attack surface and build a threat model. S4–S6 do multi-lens analysis and adversarial verification. S7–S9 dedupe, synthesize exploit chains, and emit SARIF. The new work is S10 and S11: S10 proposes fixes keyed to CWE–language–framework combinations, and S11 runs an adversarial panel that scores each fix against weighted gates and returns validated, validation_failed, or needs_review. Remediation now works with OpenAI-compatible and open-weight models, not just one vendor.

Here is the claim to be careful with. Coverage of this release has framed it as an AI that patches production before a human looks. Visa's README says the opposite: findings and fixes are "LLM-generated triage candidates that require human review." VentureBeat's own reporting has Visa saying patches land in a working copy and that nothing bypasses your normal build, test, and review flow. We are going with the primary source. Also worth noting: the repo states precision and recall figures are not yet published.

Why an open-source security harness matters for your business

You are not going to run an eleven-stage agentic pipeline against your Rails app next week. That is fine. The transferable part is the shape.

Stage 11 is the piece most teams skip. Everybody has wired an LLM into "suggest a fix." Almost nobody has wired the second agent whose only job is to argue the fix is wrong, against explicit gates, with a verdict that can come back needs_review. A generator without an adversary is a very confident intern. Adding the critic is a weekend of work and it is the difference between a patch pipeline and a patch firehose.

Second: needs_review is a first-class outcome, not a failure. Build your automation so it can say "I don't know" and route to a person. Systems that only emit pass/fail will emit the wrong one.

Sources: Visa, visa/visa-vulnerability-agentic-harness on GitHub, VentureBeat.