Visa VVAH AI Patches Code Before Human Review
Visa’s open-source VVAH finds, patches and adversarially validates code vulnerabilities by default, while humans retain review, testing and merge control.
Summary
On August 27, 2026, Visa released the open-source Visa Vulnerability Agentic Harness, or VVAH, with remediation enabled by default. A standard scan runs 11 stages and edits the target repository unless stopped after stage 9. Stage 10 writes candidate fixes to a working copy; read-only stage 11 uses an adversarial panel to return validated, failed or needs review, iterating when patches miss the exploit. VVAH does not compile, build or test patches; people authorize runs, review changes and approve merges.
Born in Anthropic’s Project Glasswing, where Claude Mythos chained minor weaknesses into exploits on Visa’s network handling billions of daily transactions, VVAH rose from 595 GitHub stars and 97 forks on July 20 to more than 2,300 stars and 300 forks by August 25, with about 9% clones per visitor. Visa runs it on its own code, though the repository accepts no external code contributions. The update adds abstract syntax tree call graphs, real-time views, MTTA observability and configurable routing among Mythos, Claude Opus, newer GPT releases and open-weight models. OpenAI-compatible backends now support remediation, while Anthropic remains the default.
Visa says Mean Time to Adapt can shrink attack-path resolution from weeks to hours, though its definitions also cover inventory freshness, exploitable paths, validation and production fixes. The launch follows Tenet Security’s GhostJacking DEF CON 34 demonstration by 18 days. Exabeam’s Steve Wilson, an OWASP LLM co-lead, favors authorization before action. Visa Consulting & Analytics, led by Carl Rutstein, adds workshops, NIST 1 to 5 assessments and roadmaps. Visa will contribute VVAH to Nvidia’s Open Secure AI Alliance and join IBM and Red Hat’s $5 billion Project Lightwell. In Q2 2026, 59% of enterprises planned to adopt or switch agent-security tools within a year; 82% mainly used provider-native controls.
Positives
- 11 automated stages take VVAH from vulnerability discovery through candidate remediation and adversarial validation.
- Attack-path resolution can fall from weeks to hours under Visa’s Mean Time to Adapt measurement.
- 2,300 GitHub stars and 300 forks by August 25 show rapid growth since the June release.
- OpenAI-compatible and open-weight models now support remediation and validation through a model-agnostic runtime.
- Three human gates preserve operator authorization, patch review and final merge approval.
- Nvidia, IBM and Red Hat collaborations could broaden VVAH’s role in securing open-source software.
Risks & concerns
- Stage 10 edits repository files before a human reviews the proposed patch.
- Stage 11 neither builds nor tests patched code, leaving exploit retesting and software validation to adopters.
- GhostJacking showed an agent changing DNS after reading a malicious payload from a log file.
- VVAH runs with elevated privileges, making repository limits, scoped credentials and restricted network access essential.
- Three different MTTA definitions could produce inconsistent executive reporting about remediation performance.
- The repository does not accept external code contributions despite VVAH being open source.