Trusys – LLM Security Scanner (VS Code Extension)

VS Code extension that runs the LLM Security Scanner inside the editor: it shows AI/LLM security findings in the Problems panel and supports optional AI analysis and database upload.

What the extension does

• Scans your code for LLM-related vulnerabilities (prompt injection, code/command injection, insecure output handling, MCP tool misuse, etc.) using Semgrep rules.
• Shows results in VS Code as diagnostics (Problems panel, inline in the editor).
• Can optionally use AI to filter false positives and improve remediation text, and upload results to your backend.

Requirements: VS Code 1.74.0+, Python 3.11+. No manual setup: the extension installs the scanner and dependencies automatically when you open a folder or run a scan. You don’t need to run any commands or configure anything.

Supported frameworks

The extension uses the same Semgrep rules as the CLI. Supported languages and LLM/framework rule sets include:

Generate Eval Tests supports FastMCP (@mcp.tool), LangChain (@tool), LlamaIndex (FunctionTool.from_defaults), and LangGraph (StateGraph with ToolNode); set Eval Framework in settings to choose. For full rule coverage and framework details, see the main repository README and rules README.

Capabilities

How to use

1. Install and run a scan

1. Install the extension (from VSIX or marketplace).
2. Open a folder (workspace) that contains your code.
3. The extension sets up the scanner automatically (you may see a short “Setting up…” or “Installing…” message the first time). No commands or settings are required.
4. Scan:
   • Automatic: Save or open a file (if Scan on Save / Scan on Open are on in settings).
   • Manual: Ctrl+Shift+P / Cmd+Shift+P → LLM Security: Scan Workspace or LLM Security: Scan Current File.
5. Open Problems (Ctrl+Shift+M / Cmd+Shift+M) to see findings; click a finding to go to the code.

2. Change what gets scanned and shown

• Severity: In settings, set llmSecurityScanner.severityFilter (e.g. ["critical", "high", "medium"]).
• Files: Use llmSecurityScanner.includePatterns and llmSecurityScanner.excludePatterns (e.g. exclude tests/, **/__pycache__/).
• Rules: Set llmSecurityScanner.rulesDirectory (default uses the rules bundled with the scanner, e.g. llm_scan/rules/python).
• Turn off auto-scan: Set llmSecurityScanner.scanOnSave and/or llmSecurityScanner.scanOnOpen to false.

3. Use AI analysis (optional)

• What it does: Sends findings to OpenAI or Anthropic to reduce false positives and improve remediation text.
• How to enable:
  1. Settings → search “LLM Security”.
  2. Enable AI Analysis and set AI Provider (e.g. openai) and AI Model (e.g. gpt-4).
  3. Set AI API Key or use env var OPENAI_API_KEY / ANTHROPIC_API_KEY.
• When it runs: During any scan (workspace or current file) when AI analysis is enabled.
• Cost: Uses the provider’s API; you can set AI Max Findings to cap how many findings are sent.

4. Upload results to your backend (optional)

• What it does: Sends scan results to your own API (e.g. for dashboards or history).
• How to use:
  1. Configure in settings: Upload Endpoint, Application ID, API Key.
  2. Run LLM Security: Scan and Upload to Database from the Command Palette (Ctrl+Shift+P / Cmd+Shift+P).
• Note: Normal Scan Workspace / Scan Current File do not upload; only the “Scan and Upload to Database” command does.

5. Generate Eval Tests (optional)

• What it does: Extracts tools from your Python code (FastMCP, LangChain, LlamaIndex, or LangGraph), calls the configured AI provider to generate test prompts with a mix of eval types (tool_selection, safety, prompt_injection, argument_correctness, robustness), and writes a JSON file (e.g. eval_tests.json). For LangGraph, the JSON includes graph_structure so you can run valid-path evals. Use the CLI to run concrete evals: python -m llm_scan.eval --eval-json <path> --graph <module:attr>.
• How to use:
  1. Set AI Provider and AI Model in settings (e.g. openai, gpt-4). Set AI API Key or use OPENAI_API_KEY / ANTHROPIC_API_KEY in the environment.
  2. Set Eval Framework to mcp (FastMCP @mcp.tool), langchain (LangChain @tool), llamaindex (LlamaIndex FunctionTool.from_defaults), or langgraph (LangGraph StateGraph with ToolNode) in settings, or choose the framework when you run the command (quick pick).
  3. Run LLM Security: Generate Eval Tests from the Command Palette (Ctrl+Shift+P / Cmd+Shift+P). You can pick FastMCP, LangChain, LlamaIndex, or LangGraph for that run (or use the current setting).
  4. Choose where to save the JSON (default: eval_tests.json in the workspace root). When it finishes, you can open the file from the notification.
• Note: Requires a workspace folder and Python files with the chosen framework’s tools. To run evals (tool-selection accuracy, valid path rate, tool coverage), use the scanner CLI: see TEST_GENERATION.md.

6. Clear results

• LLM Security: Clear Results removes all current findings from the Problems panel.

7. Reinstall or fix scanner/dependencies

• LLM Security: Install Dependencies triggers the extension’s dependency install (scanner/Semgrep) again. Use if the scanner is missing or broken.

Commands (Command Palette: Ctrl+Shift+P / Cmd+Shift+P)

Settings (search “LLM Security” in VS Code Settings)

Scan behavior

• llmSecurityScanner.enabled – Turn the extension on/off.
• llmSecurityScanner.scanOnSave – Scan when a file is saved.
• llmSecurityScanner.scanOnOpen – Scan when a file is opened.
• llmSecurityScanner.scanDelay – Delay (ms) before running a scan after a change.
• llmSecurityScanner.severityFilter – Which severities to show (e.g. ["critical","high","medium"]).
• llmSecurityScanner.includePatterns – Glob patterns for files to include (e.g. ["*.py"]).
• llmSecurityScanner.excludePatterns – Glob patterns to exclude (e.g. ["**/__pycache__/**"]).

Scanner

• llmSecurityScanner.pythonPath – Python executable used to run the scanner (e.g. python3 or venv/bin/python).
• llmSecurityScanner.rulesDirectory – Path to rules (relative to workspace or absolute).
• llmSecurityScanner.autoInstallDependencies – Whether to auto-install scanner/Semgrep on activation.

AI analysis and Eval Tests (optional)

• llmSecurityScanner.enableAiAnalysis – Enable/disable AI analysis for scans (when implemented).
• llmSecurityScanner.aiProvider – openai or anthropic. Used by Generate Eval Tests and AI analysis.
• llmSecurityScanner.aiModel – e.g. gpt-4, gpt-3.5-turbo, claude-3-opus-20240229. Used by Generate Eval Tests and AI analysis.
• llmSecurityScanner.aiApiKey – API key (or use OPENAI_API_KEY / ANTHROPIC_API_KEY). Used by Generate Eval Tests and AI analysis.
• llmSecurityScanner.evalTestMaxPromptsPerTool – Max prompts per tool when generating eval tests (default: 3).
• llmSecurityScanner.evalFramework – Framework for eval extraction: mcp (FastMCP), langchain (LangChain), llamaindex (LlamaIndex), or langgraph (LangGraph) (default: mcp).
• llmSecurityScanner.aiConfidenceThreshold – Minimum confidence for AI verdict (0–1).
• llmSecurityScanner.aiMaxFindings – Max number of findings to send to AI (limits cost).

Database upload (optional)

• llmSecurityScanner.uploadEndpoint – Backend URL (e.g. https://api.example.com/api/v1/scans).
• llmSecurityScanner.applicationId – Application ID in your backend.
• llmSecurityScanner.apiKey – API key for the upload endpoint.

Viewing results in VS Code

• Problems panel: View → Problems, or Ctrl+Shift+M / Cmd+Shift+M. Each finding shows rule, message, file, and line.
• In editor: Red/yellow/blue squiggles and gutter markers (by severity). Click to go to the line.
• Remediation: Shown in the problem message or in the hover where supported.

Troubleshooting (extension-only)

Extension doesn’t run or “scanner not found”

• Ensure Python 3.11+ is installed and that llmSecurityScanner.pythonPath points to it.
• Run LLM Security: Install Dependencies or install manually: pip install trusys-llm-scan (and ensure Semgrep is available).
• Check Output → “LLM Security Scanner” for errors.

"llm_scan package not found" or setup fails

• The extension normally sets up the scanner automatically when you run a scan. If you see this error, ensure a folder is open (File → Open Folder) and run Scan Workspace or Scan Current File again; setup will run and the scan will retry.
• If it still fails, run LLM Security: Install Dependencies from the Command Palette. You can also set llmSecurityScanner.pythonPath to a Python that already has the scanner (e.g. a venv with pip install trusys-llm-scan).

No findings

• Confirm severityFilter includes the severities you expect.
• Check includePatterns / excludePatterns (e.g. file might be excluded).
• Run Scan Workspace or Scan Current File manually and watch Output for scanner output.

AI analysis not running

• Ensure enableAiAnalysis is true and aiProvider / aiModel are set.
• Set aiApiKey or OPENAI_API_KEY / ANTHROPIC_API_KEY. Check Output for API errors.

Generate Eval Tests fails

• Set aiProvider and aiModel in settings; set aiApiKey or OPENAI_API_KEY / ANTHROPIC_API_KEY in the environment.
• Ensure the workspace has Python files that define tools for the selected Eval Framework (e.g. FastMCP @mcp.tool(), LangChain @tool, LlamaIndex FunctionTool.from_defaults, LangGraph StateGraph + ToolNode). Check Output → “LLM Security Scanner” for API or timeout errors.
• To run evals on the generated JSON, use the CLI: python -m llm_scan.eval --eval-json <path> --graph <module:attribute> (see main repo TEST_GENERATION.md).

Database upload fails

• Confirm uploadEndpoint, applicationId, and apiKey are set and that the backend is reachable.
• Use Scan and Upload to Database (upload is not done by the normal scan commands).

Performance

• Increase scanDelay, or disable scanOnSave / scanOnOpen and scan only via commands.
• Narrow includePatterns or add excludePatterns; set aiMaxFindings if using AI.

Installing the extension

From VSIX (e.g. local build)

cd vscode-extension
npm install
npm run compile
vsce package

Then in VS Code: Extensions → … → Install from VSIX → choose the .vsix file.

Development (F5)

Open the vscode-extension folder in VS Code, press F5 to launch Extension Development Host. Use llmSecurityScanner.pythonPath in the host so it uses a Python with the scanner installed.

More information

• Scanner and rules: see the main repository README and TEST_GENERATION.md for CLI, eval test generation, eval types (tool_selection, safety, prompt_injection, etc.), and running concrete evals (python -m llm_scan.eval).
• Backend/dashboard: see the main repo’s backend/ documentation for server setup.