AI security guardrail · 2026
ReachGate
Checks whether a scanner finding is actually reachable from your application's entry points. Industry data puts roughly four out of five critical findings down as noise; ReachGate turns that question into deterministic verdicts that export as OpenVEX and SARIF, re-checkable offline with no token and no network.
Solo · built for the GitLab Transcend hackathon, 2026
snapshot
- type
- AI security guardrail
- period
- 2026
- source
- public github
- demo
- watch
- pipelines
- merge request 2 · merge request 3
problem
Security scanners can mark many findings as critical without proving that the vulnerable code is reachable. ReachGate turns that question into an auditable graph search with deterministic verdicts.
outcomes
- Proved one live GitLab docs-site critical reachable and one not reachable, each with a re-checkable receipt
- Turned noisy scanner criticals into deterministic, re-checkable verdicts
- Kept verdicts auditable: an sha256 manifest over the exported artifacts by default, Ed25519 signing as an opt-in extra, both verifiable offline
what i built
- Bounded BFS over GitLab Orbit code graph data
- Reachability verdicts for vulnerable code paths
- Deterministic rule engine where AI only explains the result
- UNKNOWN verdict when a search exceeds its budget
- OpenVEX and SARIF exports for security tooling
- sha256 manifest generation
- Optional Ed25519 signature support
- Offline verifier that needs no token and no network
- CLI workflow for local checks
- GitLab CI merge request bot with duplicate-comment prevention
- GitLab Duo AI agent integration over the Orbit MCP server
- A build-breaking test that rejects any fabricated 'safe' verdict
tech stack
PythonGitLab Orbit graph APIBFSGitLab CI/CDOpenVEXSARIFEd25519MCPpytest
Read the source
The repo has the code, structure, and implementation detail.