Visa Vulnerability Agentic Harness (VVAH) is Visa's harness for finding, fixing, and validating security vulnerabilities in a codebase with large language models, built on learnings from Anthropic's Project Glasswing. It is designed around triage speed rather than discovery, and its main metric is Mean Time to Adapt: the time from an AI-discovered exploitable weakness to a validated fix in production.
A scan runs an optional static seed (S0) and up to eleven stages in four phases. Discovery and modeling (S1–S3) maps the attack surface and builds a threat-aware plan. Deep dive and verification (S4–S6) applies multiple analysis lenses and adversarial verification. Synthesis and reporting (S7–S9) deduplicates and chains findings into Markdown and SARIF 2.1.0 reports with CVSS and CWE. Optional remediation and validation (S10–S11) propose fixes and adversarially grade them, and are off in the default profile because they edit source files.
Threat modeling before analysis narrows the attack surface, multi-agent deterministic voting reduces false positives, AST and call-graph seeding focuses review on reachable code, and interprocedural taint analysis adds data-flow evidence. Every model-driven stage can use its own model and provider: Anthropic Claude natively, OpenAI or any OpenAI-compatible gateway, or open-weight models on Chat Completions-compatible endpoints, with routes that include the Claude Code CLI and Deep Agents.
A beta exploit-verification step attacks API findings with live HTTP traffic against a target running on your own machine, using a supplied Postman or OpenAPI collection, and a replay command re-runs the confirmed exploit after a fix ships. Findings and fixes are LLM-generated triage candidates that need human review. VVAH does not build, run, or test the target, publishes no accuracy numbers yet, and has no global token spend cap.
Features
- Staged pipeline: threat-modeled discovery, multi-lens deep dive, adversarial verification, reporting, remediation, and fix validation (S0–S11)
- False-positive reduction: multi-agent deterministic voting and adversarial verification of candidate findings
- Reachable-code focus: AST and call-graph seeding plus interprocedural taint analysis
- Structured output: Markdown and SARIF 2.1.0 reports with CVSS, CWE, findings JSON, and per-run diagnostics
- Exploit verification (beta): confirms API findings with live HTTP requests to a local target and replays the exploit after a fix
- Remediation and validation: optional candidate fixes, report-only or applied, graded by adversarial validation
- Per-role models: each stage can use Anthropic, OpenAI, an OpenAI-compatible gateway, or an open-weight endpoint
- Portfolio scans: CSV-defined repository lists with resumable state and progress monitoring
- Coverage backstop: unrecognized files go into catch-all review chunks rather than being skipped
- Agent setup: writes instruction files and a skill for Claude Code, GitHub Copilot, and Gemini CLI
- Cost estimate: an estimate command sizes token spend before a scan
