Correctover AI Security Scanner — GitHub Action

Scan your AI agent code for command injection, path traversal, SSRF and 19 categories of AI-specific vulnerabilities. Runs in seconds.

Quick Start

name: AI Security Check
on: [push, pull_request]

jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      
      - name: Run Correctover Security Scan
        uses: Correctover/correctover-action@v1
        with:
          level: L2
          fail-on: critical

That's it. Your CI will now catch AI-specific security issues on every push.

What It Detects

Category	Severity	Examples
Command Injection	CRITICAL	os.system(), subprocess(shell=True), eval()
Path Traversal	CRITICAL	open(user_path) without validation
SSRF	CRITICAL	fetch(user_url) without allowlist
SQL Injection	CRITICAL	f"SELECT ... {user_input}"
Code Injection	CRITICAL	exec(user_input)
Hardcoded Secrets	HIGH	API keys in source
Unsafe Deserialization	HIGH	pickle.loads()
+ 12 more categories

Inputs

Input	Default	Description
path	.	Path to scan
level	L2	L1 (fast syntax) or L2 (contextual analysis)
fail-on	critical	Fail CI on: critical, high, or medium
output-format	json	terminal, json, markdown
save-output	correctover-results.json	Output file path

Outputs

Output	Description
findings-count	Total findings
critical-count	CRITICAL findings
high-count	HIGH findings
baseline-status	PASSED or FAILED

Advanced Usage

Fail on HIGH+ findings

- uses: Correctover/correctover-action@v1
  with:
    fail-on: high
    level: L2

Scan a specific directory

- uses: Correctover/correctover-action@v1
  with:
    path: ./mcp-servers/
    level: L2

Use results in later steps

- uses: Correctover/correctover-action@v1
  id: scan
  with:
    output-format: json
    save-output: results.json

- name: Check results
  if: steps.scan.outputs.baseline-status == 'FAILED'
  run: echo "Security issues found!"

Real-World Validation

We blind-tested this scanner on 27 MCP Server projects from GitHub:

• 6 top projects (Cline, OpenHands, Aider, Continue, Goose, Authgear) — all passed ✅
• 21 mid-tier projects — found 1 confirmed CRITICAL vulnerability (CVSS 9.8)
• False positive rate: 3.8% at CRITICAL+HIGH level

Why AI-Specific?

Traditional SAST tools (SonarQube, Semgrep, Snyk) don't understand:

• MCP tool call patterns
• Agent trust chains
• LLM-generated command execution
• AI framework-specific safe/unsafe API usage

Correctover does. It's built specifically for the AI agent code supply chain.

CCS Compliance

Every finding maps to the CCS Runtime Verification Standard — a 6-dimension framework covering Structure, Schema, Latency, Cost, Identity, and Integrity verification.

CCS has been submitted to IETF (draft-correctover-ccs-00).

Links

• Scanner source
• CCS Standard
• Correctover.com
• Community Discussions

License

Apache 2.0