LLMScan - AI Security Scanner

English | 日本語 (Japanese)

Local Static Application Security Testing (SAST) for AI & Web Applications

LLMScan analyzes Python, TypeScript, JavaScript, and PHP source code in real time as you edit. It traces data flow from untrusted sources to dangerous sinks, detecting both traditional web vulnerabilities (OWASP Top 10) and LLM-specific security threats (OWASP LLM Top 10 - LLM05: Insecure Output Handling).

All scans run 100% locally inside your editor using Tree-sitter WebAssembly. Zero code is sent to external servers, making it fully safe for proprietary codebases and untrusted workspaces.

🌟 Key Features

• Injection Vulnerability Detection (OWASP A05): SQL Injection (Raw ORM & parameterized queries), Command Injection (exec, spawn, shell=True), Code Injection (eval, new Function), Insecure Deserialization (pickle, unserialize, yaml.unsafe_load), Cross-Site Scripting (XSS), Server-Side Template Injection (SSTI), and Path Traversal / File Inclusion.
• OWASP LLM05 (Insecure Output Handling) & Agentic Security: Tracks responses from OpenAI, Anthropic, Gemini, Ollama, LangChain, LlamaIndex, and Vercel AI SDK, as well as AI Agent tool calls (tool_call.function.arguments, tool_input) and Model Context Protocol (MCP) tool handlers (request.params.arguments) directly to dangerous sinks (innerHTML, sqlite3.execute, subprocess.run, eval).
• Interactive Tree View & Activity Bar: Dedicated Activity Bar icon (llm.svg) and sidebar view with severity badges, line-by-line navigation, and a quick toggle filter for High/Critical issues.
• Intelligent Taint Tracking & Sanitizers: Built-in understanding of safe sanitizers: Zod schema validation, Yup, Pydantic, cryptographic hash digests (hashlib, crypto.createHash, bin2hex), and type coercion (+x, Number(), int(), filter_var).
• Actionable Hovers & QuickFix: Hover over any wavy line to see the vulnerability explanation, language-specific remediation examples, complete taint propagation trace, and OWASP/CWE references. Insert reason-enforced suppression comments with one click (QuickFix).
• Baseline Support: Create .llmscan-baseline.json to suppress legacy technical debt and only alert on newly introduced vulnerabilities.

📖 Usage

⚙️ Configuration

You can configure options in .llmscan.yml at the repository root (shared with the CI CLI) or via VS Code Settings:

# .llmscan.yml
minConfidence: medium        # low | medium | high
failOn: high                 # Failure threshold for CI CLI
baseline: .llmscan-baseline.json
exclude:
  - "**/tests/**"
  - "**/generated/**"
sanitizers:                  # Custom sanitization functions defined in your project
  javascript: ["mySanitizer"]
  python: ["app.security.sanitize_query"]
  php: ["clean_input"]

VS Code Settings (settings.json)

• llmscan.enable: Enable or disable LLMScan security analysis. (Default: true)
• llmscan.minConfidence: Filter findings by minimum confidence level (low, medium, high).
• llmscan.showBaselined: Whether to display existing findings recorded in the baseline file. (Default: false)

🛡️ Supported Rules

🔒 Privacy & Local Processing

LLMScan is built on local Tree-sitter WebAssembly grammar parsers. No source code, file paths, or telemetry are ever transmitted across the network. It functions completely offline and in air-gapped environments.