Safety-first architecture
- establishedevidence level
- architecture
- 2026-11-15
- #vulnmind
- #safety
- #architecture
Vulnmind runs live tests against real systems, so the first design question was how it should refuse to act, not what it can do.
Fail-closed scope guards
Before anything runs, the tool checks that the API permissions it has been given look the way they should. If they look wrong, it stops. It does not warn and continue. A guard that fails open is a guard that fails on the one day it matters.
Rate limiting
Request rate is capped in the tool itself, so a bug or a bad loop cannot turn into a flood of traffic against someone else’s system.
Approval-gated dispatch
Every live test goes through a generated exploration plan. The plan carries a cryptographic fingerprint. Nothing is dispatched until that exact plan has been approved, so what runs is provably what was reviewed.
Why this order
Capability is easy to add later. Safety constraints bolted on afterwards tend to have gaps where the capability was already assumed. Building the refusal paths first means every new capability has to fit inside them.