---

The Problem

AI accelerates code generation but creates new security bottlenecks:

Issue	Impact
Security reviews block MRs for hours or days	7 hrs/week lost per team member
AI makes compliance management harder	70% of teams report this
Compliance issues discovered after deployment	76% of organizations
Traditional SAST tools miss AI-specific risks	Prompt injection, LLM output-to-exec, unsafe ML deser
40-62% of AI-generated code has vulnerabilities	No existing tool detects AI-specific threat patterns

_{Source: GitLab 2025 Global DevSecOps Report}

The Solution

AgentFlow Auditor — Four AI agents that catch security risks traditional SAST misses in AI-generated code. 41 detection rules including LLM prompt injection, output-to-exec, unsafe deserialization, SQL injection, SSRF, and path traversal. Posts scannable reports, generates fix patches, tracks risk drift. One @mention triggers everything.

---

Architecture

  Developer opens MR
        │
        ▼
  @duo-agentflow-auditor review this MR
        │
        ▼
┌───────────────────────────────────────────────────────┐
│                                                       │
│   ┌─────────────┐       ┌──────────────────┐         │
│   │   Scanner   │──────▶│    Reporter      │         │
│   │   Agent     │       │    Agent         │         │
│   │             │       │                  │         │
│   │ Read diffs  │       │ Grade risk       │         │
│   │ Match 34    │       │ Risk heatmap     │         │
│   │ rules       │       │ Post MR comment  │         │
│   │ Score risk  │       │ Create issue     │         │
│   └─────────────┘       └────────┬─────────┘         │
│                                  │                   │
│                          ┌───────┴────────┐          │
│                          │  SAFE?         │          │
│                          │  ├─ Yes ──────────────┐   │
│                          │  └─ No ───┐   │       │   │
│                          └───────────┘   │       │   │
│                                  │       │       │   │
│   ┌─────────────┐       ┌────────▼───────┘┐      │   │
│   │   Metrics   │◀──────│    Fixer        │      │   │
│   │   Agent     │       │    Agent        │      │   │
│   │             │◀──────────────────────────────┘   │
│   │ Baseline    │       │ Confidence-scored│         │
│   │ Cross-MR    │       │ patches         │         │
│   │ Green       │       │ Fix branch + MR │         │
│   │ Posture     │       │                 │         │
│   └─────────────┘       └─────────────────┘         │
│                                                       │
│            GitLab Duo Agent Platform (ambient)        │
│                                                       │
│   ┌───────────────────────────────────────────┐       │
│   │  External SAST Agent (CI/CD container)    │       │
│   │  bandit + semgrep + custom rules merge    │       │
│   └───────────────────────────────────────────┘       │
└───────────────────────────────────────────────────────┘

Agent Roster

Agent	Role	Tools	Key Capability
Scanner	Analyze MR diffs	10 tools	41 rules (26 regex + 15 Semgrep), AI-specific threat detection, vulnerability integration
Reporter	Post audit reports	7 tools	Scannable in 10s (grade + heatmap + top 5), vulnerability linking, auto issue on DANGER
Fixer	Generate code fixes	8 tools	Confidence-scored patches (HIGH/MEDIUM/LOW), auto fix MR creation
Metrics	Track risk baseline	6 tools	Cross-MR learning, team posture, baseline drift, energy/carbon tracking
SAST Scanner	External SAST	CI/CD	Runs bandit + semgrep, merges with custom rules via Python script

---

Features

Security Scanning 41 Detection Rules — 26 regex + 15 Semgrep custom rules 8 Risk Categories — From destructive commands to prompt injection AI-Specific Detection — LLM prompt injection, output-to-exec, unsafe deserialization Risk Scoring — 0-100 per finding (severity x context x category) Grade System — SAFE / WARNING / DANGER	Automation One-trigger activation — @mention or assign reviewer Conditional routing — SAFE scans skip fixer, saving tokens Structured MR comments — Risk tables, fix suggestions, collapsible details Auto-fix generation — Code patches on a new branch Issue creation — Automatic on DANGER grade
Baseline Tracking Risk drift detection — Compare scans over time Trend analysis — Improving / degrading / stable Fix adoption rate — Track how many suggestions were applied History log — JSONL append per scan	Green Metrics Token tracking — Usage per scan Energy estimation — kWh per scan Carbon footprint — kg CO2 with real-world analogies Optimization suggestions — Reduce scan scope, cache baselines

---

Detection Categories

Semgrep Custom Rules

15 production-grade Semgrep rules across 6 categories: