Security Guardian: Autonomous DevSecOps Flow

Autonomous AI agent auditing Merge Requests for security flaws. It moves beyond chat, identifying vulnerabilities like SQLi or Secrets and suggesting secure code patches directly in the workflow.

Comment

Inspiration

Every 39 seconds a company gets hacked. Most startups ship vulnerable code because they simply cannot afford a dedicated security engineer. We watched teams merge SQL injection, hardcoded AWS keys, and broken authentication straight to production — not because they were careless, but because security review is expensive and slow.

We asked: What if every developer had an AI security engineer that never sleeps, never misses a PR, and catches vulnerabilities before they reach production?

That became Security Guardian.

What it does

Security Guardian is a fully autonomous DevSecOps agent powered by Claude (Anthropic) via the GitLab Duo Agent Platform. It automatically triggers on every Merge Request and executes an 8-phase security review:

  1. Reconnaissance — Reads all changed files, scans the entire repo for secret patterns
  2. Taint Analysis — Traces user input from Source → Sink to find injection paths like a senior security researcher
  3. Vulnerability Detection — Detects 14+ vulnerability types including SQL Injection, XSS, hardcoded secrets, broken auth, insecure deserialization, IDOR, command injection, path traversal, weak crypto, and dependency CVEs
  4. Auto-Remediation — Rewrites vulnerable code with before/after explanation so developers learn
  5. Issue Tracking — Creates detailed GitLab issues with real breach examples and dollar costs
  6. Compliance Mapping — Maps every finding to GDPR, PCI-DSS, HIPAA, and SOC2
  7. Audit Report — Posts Security Score 0-100 with APPROVE / REQUEST CHANGES / BLOCK verdict
  8. Merge Blocking — Adds security-review-required label to block unsafe merges automatically

Demo results on real code:

  • Found 17 vulnerabilities including $2.7M–$410M in potential breach exposure
  • Auto-fixed all critical vulnerabilities
  • Security Score improved from 0/100 → 95/100
  • Tested on real OWASP WebGoat code — found 16 vulnerabilities across SQL Injection, XSS, XXE, SSRF, CSRF, Path Traversal, and Insecure Deserialization

How we built it

  • Custom Agent (agents/agent.yml) — 7-phase system prompt with taint analysis logic, 45 GitLab tools configured, compliance mapping for GDPR/PCI-DSS/HIPAA/SOC2
  • Custom Flow (flows/flow.yml) — Auto-triggers on MR assignment or @mention, extracts MR IID from context, executes full 8-step audit autonomously
  • Claude (Anthropic) — Powers all reasoning, taint analysis, and vulnerability detection via GitLab Duo Agent Platform
  • GitLab Duo Agent Platform — Provides tool access: get_merge_request, edit_file, create_issue, create_merge_request_note, update_merge_request and 40+ more
  • Green Code — Detects O(n²) loops, N+1 queries, estimates CPU and energy savings per fix

Challenges we ran into

  • Goal too long error — GitLab passes the full MR diff as goal context; large repos like WebGoat hit the 16384 character limit. Fixed by optimizing the system prompt size.
  • Windows CRLF line endings — The edit_file tool uses Unix LF for matching; Windows files caused all edits to fail silently. Fixed by instructing the agent to fall back to create_file_with_contents.
  • GraphQL 422 errors — Em-dashes — and pipe characters | in YAML caused GraphQL failures. Fixed by replacing all special characters.
  • MR context extraction — The flow initially received just a number as goal. Fixed by parsing MergeRequest IID: X from the context string.
  • Agent trying to delete files — Without explicit rules, the agent tried deleting and recreating files when edits failed. Fixed with strict golden rules in the prompt.

Accomplishments that we're proud of

  • Fully autonomous — Zero human involvement from MR creation to BLOCK/APPROVE verdict
  • Real OWASP testing — Successfully scanned real WebGoat vulnerable code, not just toy examples
  • $410M+ breach exposure detected in a single scan
  • Complete compliance coverage — GDPR, PCI-DSS, HIPAA, SOC2 mapped automatically
  • Green Code — Detected O(n²) loops with 99% CPU reduction and ~1,800 kWh/year energy savings estimated
  • Security Score progression — 0/100 → 95/100 in a single MR review cycle
  • 14 GitLab issues created automatically with CWE references, breach examples, and remediation steps

What we learned

  • AI agents can replace entire security workflows when given the right tools, context, and chain-of-thought reasoning
  • Prompt engineering is engineering — every character counts when you hit API limits
  • Taint analysis in natural language works surprisingly well when structured as Source → Sink → Sanitizer
  • GitLab Duo Agent Platform is genuinely powerful — the combination of 45+ tools with Claude's reasoning creates something that feels like a real team member
  • Every dev team regardless of size or budget deserves an autonomous security engineer

What's next for Security Guardian: Autonomous DevSecOps Flow

  • Google Cloud Pub/Sub integration — Publish every vulnerability as a real-time event so subscribed teams get instant alerts
  • Auto-approve loop — Guardian re-reviews after fixes and auto-approves when score hits 100/100
  • Multi-language support — Extend beyond Java to Python, Node.js, Go, Ruby
  • Historical trending — Track security score across MRs over time
  • Slack/Teams notifications — Alert security team when Critical vulnerabilities are found
  • Custom rule engine — Let teams define their own vulnerability patterns
  • Publish to GitLab AI Catalog — Make Security Guardian available to all GitLab users worldwide

Built With

  • agent
  • antropic
  • ci/cd
  • claude
  • duo
  • gitlab
  • java
  • owasp
  • platform
  • yaml
Share this project:

Updates