CONTROL
Rule engine owns verdict
Findings, severity, category scores, safety score, attack paths, patch previews, and gate decisions are generated by backend logic.
METHODOLOGY
How the scanner decides risk.
The security verdict is deterministic. Gemini explains results only after rule-based scanning creates findings, severity, scores, patch previews, attack paths, and deployment gate decisions.
Scan architecture
Text-only source analysis pipeline.
01
Load sourceGitHub ZIP, uploaded ZIP, or demo agent files are read as text only.
→
02
Run scannersRule modules detect agent-specific risks and produce stable finding IDs.
→
03
Build artifactsThe backend generates Prove Mode paths, patch previews, and a deployment gate decision.
→
04
Explain safelyGemini summarizes the deterministic result. It does not decide the verdict.
→
05
Review reportFrontend shows score, risk panels, findings, patches, DAP, and export actions.
CONTROL
AI explains only
Gemini writes compact summaries and DAP answers grounded in the report. It does not invent findings or change scores.
CONTROL
Safe by design
GitHub repositories and ZIP files are safely extracted, limited, read as text, and never executed by the scanner.
Risk categories
Six categories, one deployment decision.
Category
What it checks
Report output
Prompt Injection
Committed prompts, unsafe prompt assembly, raw user input inside agent instructions.
Prompt Injection Risk
Secret Exposure
Hardcoded keys, tokens, JWT secrets, and credential-like source values.
Secret Exposure Risk
Tool Permission
Unscoped tools, dangerous functions, shell/code execution, external actions.
Tool Permission Risk
Human Approval
High-impact actions without confirmation, review, or approval patterns.
Human Approval Risk
Data Exposure
PII-like fields, customer records, and unmasked sensitive outputs.
Data Exposure Risk
Auditability
Missing logging around tool calls, agent decisions, or critical actions.
Auditability Risk
Scoring
Safety score vs category score.
The overall safety score is higher when the agent is safer. Category scores are risk scores, so higher means worse in that specific area.
- 0–39 category score: lower risk
- 40–69 category score: watch and review
- 70–100 category score: fix first
Deployment gate
BLOCK, REVIEW, or ALLOW.
The gate uses safety score, critical findings, secret exposure, unsafe high-impact tools, and missing approval gates to decide if a repo should move forward.
- BLOCK: unsafe before deployment
- REVIEW: proceed only after human review
- ALLOW: no hard blocker detected