The natural assumption about AI-powered security tooling is that the hard part is finding vulnerabilities. Visa's newly open-sourced VVAH (Visa Vulnerability Agentic Harness) is built around a different premise. The hardest part isn't discovery. It's triage.
What VVAH Is
VVAH is a four-phase, eleven-stage pipeline from code repository to validated fix. Built on Project Glasswing (Anthropic's initiative for AI-assisted vulnerability research), it works with Claude, OpenAI-compatible models, or both.
Primary metric: MTTA - Mean Time to Adapt (AI-discovered exploitability ? validated fix in production).
Four phases:
- Phase 1 - Discovery & Modeling (S1-S3): Attack surface mapping, STRIDE threat model, prioritized hunt plan
- Phase 2 - Deep Dive & Verification (S4-S6): Multi-lens analysis (6 specialized roles), policy gates, adversarial verification
- Phase 3 - Synthesis & Reporting (S7-S9): Deduplication, exploit chain construction, SARIF 2.1.0 emission
- Phase 4 - Remediation & Validation (S10-S11): Minimal fix proposals, adversarial panel validation
The Triage Design
The key structural choice: deterministic controls sit before expensive downstream LLM work. S5 applies policy gates; semantic dedup fires only when survivors exceed a threshold (default 25). Then S6 runs adversarial verification - a separate LLM pass constructing exploit chains and tracing trust boundaries.
Findings that survive S5+S6 are confirmed exploitable. S7-S9 chain and emit SARIF over a clean signal.
Skills as the Modularity Unit
Each LLM-driven stage is a composable, independently replaceable skill. Pass --stop-after s9 for detection only (no code edits). The default profile edits source files in your target repo via S10 fix mode - document this before onboarding teams.
Multi-Model by Design
Three backends: Anthropic Python SDK, claude CLI subprocess, OpenAI-compatible API. Detection pipeline (S1-S9) is backend-agnostic. Full remediation + validation (S10-S11) currently requires Anthropic models.
Why This Architecture Matters
The same pattern scales to any agentic pipeline needing to convert AI signals into validated outputs:
- Threat model before analysis
- Adversarial verify before synthesis
- Measurable metric (MTTA) connecting AI capability to operational velocity
- Composable skills tunable per language/framework/CWE
Repo: github.com/visa/visa-vulnerability-agentic-harness
Originally published at vsebude.it
Top comments (0)