AI Incident Response & Vulnerability Agent
An IR and vulnerability agent, built on Visa's open vulnerability harness, that does not just generate findings. It validates them.
The problem
Purpose-built security models are impressive in a demo and unreliable in an investigation. They hallucinate findings and need heavy tuning to be usable. For threat analysis and incident response, a confident wrong answer is worse than no answer, because an analyst spends real time chasing it. Optimising the model alone never closed that gap.
What I built
So I stopped optimising the model and started building around it. On Visa's open vulnerability harness I built an incident-response and vulnerability agent whose job is not to generate findings but to earn the right to surface one.
- Adversarial verification runs against each finding to confirm it is actually exploitable before it counts.
- Multiple agents have to agree on a result before it is kept, so one model's guess is not enough.
- Proposed fixes go through a review panel that grades them and returns a verdict.
The result
The unverified guesses get filtered out. What reaches a human is triage they can act on. It reframed the work for me: AI engineering for security is less about the model and more about the verification layer you wrap around it. The model generates, and the harness decides what you are allowed to believe.