Visa’s open-source Visa Vulnerability Agentic Harness (VVAH) now detects vulnerabilities, writes candidate fixes into a target repository, and runs an adversarial validation panel against those patches before any human merges them. The company says the full 11-stage loop ships by default unless an operator configures the run to detection-only.
How the flow works
A default VVAH run executes all 11 stages and — unless stopped — edits source files in the target repo. Stage 10 writes candidate fixes to a working copy, and stage 11 runs a read-only adversarial validation panel that scores each fix, returning one of three verdicts: validated, validation failed, or needs review. According to Visa’s README, stage 11 runs in read-only mode and the harness does not compile, build, or run tests against the patched tree; build/test/code-review continue to be the organization’s responsibility.
Visa paired the release with an expansion of Visa Consulting & Analytics. Rajat Taneja, Visa’s president of technology, framed the product as a response to a shifted bottleneck: "AI is finding vulnerabilities faster than humans can in the history of our technology industry. The new bottleneck is fixing and proving we have fixed things."
Origins and adoption
VVAH grew out of Visa’s participation in Anthropic’s Project Glasswing, where Claude Mythos chains were used to probe network paths behind billions of transactions; Visa initially used Mythos extensively for semantic reasoning. The harness went public on GitHub in June. Its GitHub metrics rose from 595 stars and 97 forks on July 20 to more than 2,300 stars and 300 forks by August 25, with a clone-to-visitor ratio Visa put near 9%.
Visa says it was "client zero" — the tool was designed for Visa’s own use and then released for others. The repository currently does not accept external code contributions, so the harness that edits adopters’ source does not pull code into Visa’s repo.
From discovery to validated remediation
The new release extends the pipeline beyond discovery and reporting to include remediation, validation, and iteration. Underneath the loop, the scanner was refactored around an abstract syntax tree (AST) call graph that maps subroutine calls and the traversal paths an attacker might reach. Visa says this reduces token usage while improving reasoning, context, and exploitability analysis. The release also adds MTTA observability across stages and real-time progress views.
Visa defined Mean Time to Adapt (MTTA) — a metric it introduced with the harness — more concisely in this release as the time between discovery and resolution of attack paths. Visa claims some resolutions have shrunk from weeks to hours. The Project Glasswing white paper tracks MTTA in three dimensions (inventory freshness, exploitable paths per release, and validation cycle time), and the repository also records elapsed time from AI-discovered exploitability to a validated production fix. Visa cautions that board slides may cite the shortest interval, and recommends asking for all three measures.
Debate over approval gates
Security experts pointed out risks in an automated fix flow. Steve Wilson, Chief AI and Product Officer at Exabeam and project co-lead for the OWASP Top 10 for LLM Applications, argued the default should include an authorization gate that prevents an agent from granting itself the authority to make critical infrastructure changes. In VentureBeat commentary, he noted that rules inside prompts are suggestions, not enforceable controls.
Visa answered that VVAH is intended to be run by authorized operators against code they own in controlled environments. The company emphasized that humans remain "the gate in three places": before running the tool, when reviewing patches, and before anything is merged. Visa called VVAH a "harness, not a merge tool," and reiterated that the final call on any fix remains with the security and engineering teams.
Under the shipped default, the human gates sit outside the model: operators decide to run the tool, reviewers examine patches written to working copies, and teams perform their normal build/test/code-review and merging workflows. Visa framed this as helping security teams "work at AI speed" by automating repetitive tasks—finding issues, testing whether they are real, and proposing fixes—while leaving judgment calls to humans.
What the harness automates and what it does not
The harness automates an adversarial validation step that attempts to break candidate fixes before they reach human reviewers; failed fixes feed into an automated iterate step. However, the harness does not automatically merge fixes into main branches, nor does it compile or run tests against the patched tree as part of stage 11. Visa describes this wrapping as the governance layer and emphasizes that turning automated write access on requires careful operational controls.
The README warns that the tool runs with elevated privilege and that any role routed through certain backends may send prompt data to that provider’s endpoint. Visa recommends running scans in ephemeral environments with scoped credentials, no production secrets, and limited network access — and starting with --stop-after s9 to inspect SARIF output before any write-capable run.
Multi-model orchestration and configuration
VVAH supports multi-model orchestration. Taneja explained that different models have different strengths (for example, Mythos offered high recall while another model, Opus, had high precision) and that customers can select per-stage models via configuration. Initially, applying a fix required Anthropic backends; the current README extends remediation and validation to OpenAI-compatible and open-weight models through a shared, model-agnostic runtime, with Anthropic still the default routing for the relevant stages.
Market context and consulting
VentureBeat research cited by Visa found 59% of enterprises plan to adopt or switch agent security tooling within a year, and 82% still rely primarily on provider-native controls. Visa said it will contribute VVAH to Nvidia’s Open Secure AI Alliance as a model-agnostic framework and collaborate with Project Lightwell (IBM and Red Hat’s initiative) to harden open-source components.
Visa Consulting & Analytics is adding executive workshops, a VVAH-informed maturity assessment scored on a NIST 1–5 scale, and a cyber risk prioritization roadmap. Carl Rutstein, global head of Visa Consulting & Analytics, summarized the rationale: "Finding vulnerabilities is no longer the hardest part. Speed to remediation is the new battleground."
Practical checklist before enabling write mode
Visa’s documentation suggests concrete steps organizations should take before allowing the harness to write to repos:
- Start runs with --stop-after s9 and inspect SARIF output before any run that can write to source files.
- Map the three human gates (run approval, patch review, merge) to named owners in your processes.
- Keep build/test/code-review in your team’s flow and perform an exploit re-test before merge.
- Limit which repositories the harness can reach, run in ephemeral environments with scoped credentials and no production secrets, and restrict network access.
- Assign models per stage deliberately and measure precision/recall for your context.
Conclusion
Visa’s VVAH release automates detection, proposed remediation, and an adversarial validation step in a single pipeline intended to shorten Mean Time to Adapt. The company maintains that human approval points, scoped deployment, and governance are central to safe operation, while promoting the harness as a tool to help security teams scale triage and remediation at AI speed rather than replace human decision-makers.
