TensorClad

AI-Native Application Security Scanner for VS Code

Installation Features Demo Vulnerabilities Configuration Contributing

Demo

TensorClad detecting security vulnerabilities in real-time as you code

Why TensorClad?

As AI applications become mainstream, a new class of security vulnerabilities has emerged. Traditional SAST tools excel at finding SQL injection and XSS, but they're blind to prompt injection, API key leakage in LLM configs, and unvalidated model outputs.

TensorClad fills this gap. It's a static analysis tool built specifically for developers working with OpenAI, LangChain, Anthropic, and other AI frameworks.

The Problem

Consider this typical AI application code:

# This code has 3 security issues

api_key = "sk-proj-abc123..."  # TC001: Hardcoded API key

def chat(user_input):
    prompt = f"Help the user with: {user_input}"  # TC010: Prompt injection risk
    response = openai.chat.completions.create(
        model="gpt-4",
        messages=[{"role": "user", "content": prompt}]
    )
    return response.choices[0].message.content  # TC030: Unvalidated output

TensorClad detects all three issues as you type, showing warnings in the Problems panel with explanations and fix suggestions.

Features

Real-Time Scanning

No need to run a separate CLI tool. TensorClad analyzes your code on every keystroke and highlights issues inline with squiggly underlines, just like TypeScript errors.

Security Dashboard

Run TensorClad: Show Security Report to see a summary of all detected vulnerabilities across your workspace, organized by severity and file.

Framework-Aware Detection

Purpose-built detection rules for popular AI/LLM frameworks:

Installation

From VS Code Marketplace

1. Open VS Code
2. Press Ctrl+Shift+X (Windows/Linux) or Cmd+Shift+X (Mac)
3. Search for "TensorClad"
4. Click Install

From VSIX (Manual)

code --install-extension tensorclad-0.1.0.vsix

Detected Vulnerabilities

TensorClad identifies security issues specific to AI/LLM applications. Each finding includes a code (e.g., TC001), severity level, and remediation guidance.

API Key Exposure (TC001-003)

Hardcoded API keys are the most common security issue in AI applications.

# ❌ Detected: TC001 - OpenAI API key in source code
openai.api_key = "sk-proj-abc123def456..."

# ✅ Secure: Load from environment
openai.api_key = os.getenv("OPENAI_API_KEY")

Prompt Injection (TC010)

Direct user input in prompts allows attackers to manipulate LLM behavior.

# ❌ Detected: TC010 - User input directly in prompt
prompt = f"Summarize this text: {user_input}"

# ✅ Secure: Validate and sanitize input
prompt = f"Summarize this text: {sanitize_input(user_input)}"

Hardcoded System Prompts (TC020)

System prompts in source code can leak business logic and are hard to update.

# ⚠️ Warning: TC020 - Consider externalizing prompts
messages = [
    {"role": "system", "content": "You are a helpful assistant..."}
]

# ✅ Better: Load from configuration
messages = [
    {"role": "system", "content": load_prompt("assistant")}
]

Unvalidated LLM Output (TC030)

LLM responses are untrusted. Never execute them directly.

# ❌ Detected: TC030 - Executing unvalidated output
result = response.choices[0].message.content
exec(result)  # Remote code execution risk!

# ✅ Secure: Validate output before use
result = response.choices[0].message.content
if is_safe_output(result):
    process(result)

PII Leakage (TC050)

Logging user data can violate privacy regulations.

# ❌ Detected: TC050 - PII in logs
print(f"User email: {user.email}, Query: {query}")

# ✅ Secure: Redact sensitive data
print(f"User: [REDACTED], Query: {redact_pii(query)}")

Insecure Tool Execution (TC060)

AI agents that execute arbitrary code need strict validation.

# ❌ Detected: TC060 - Dynamic code execution
eval(llm_response)

# ✅ Secure: Whitelist allowed operations
if operation in ALLOWED_OPERATIONS:
    execute_sandboxed(operation)

Complete Rule Reference

Commands

Open the Command Palette (Ctrl+Shift+P) and type "TensorClad":

Configuration

Customize TensorClad in your VS Code settings (settings.json):

{
  "tensorclad.enabled": true,
  "tensorclad.scanOnSave": true,
  "tensorclad.scanOnOpen": true,
  "tensorclad.excludePatterns": [
    "**/node_modules/**",
    "**/dist/**",
    "**/.venv/**"
  ],
  "tensorclad.gitHooks.enabled": true,
  "tensorclad.gitHooks.blockOnError": true,
  "tensorclad.gitHooks.blockOnWarning": false
}

Configuration Options

Git Hooks Integration

TensorClad can integrate with Git to prevent pushing code with security vulnerabilities. This provides a last line of defense before insecure code reaches your repository.

How It Works

When installed, TensorClad adds pre-commit and pre-push hooks that scan staged/changed files for security issues:

• Pre-commit hook: Scans staged files before each commit
• Pre-push hook: Scans all changed files before pushing to remote

If critical vulnerabilities are found (based on your configuration), the commit or push is blocked with a detailed report.

Installing Git Hooks

1. Open the Command Palette (Ctrl+Shift+P)
2. Run TensorClad: Install Git Hooks
3. The hooks will be installed to your repository's .git/hooks/ directory

What Gets Blocked

By default, pushes are blocked when code contains:

You can configure whether warnings (non-error severity issues) also block pushes via tensorclad.gitHooks.blockOnWarning.

Bypassing Hooks (Emergency Only)

If you need to push despite warnings (not recommended):

git push --no-verify

⚠️ Warning: This bypasses all security checks. Use only when absolutely necessary.

Supported Languages

Contributing

Contributions are welcome! Whether it's adding new detection rules, improving documentation, or fixing bugs.

Quick Start

# Clone the repository
git clone https://github.com/santhoshravindran7/TensorClad.git
cd TensorClad

# Install dependencies
npm install

# Compile
npm run compile

# Watch mode (auto-recompile)
npm run watch

# Launch extension in debug mode
# Press F5 in VS Code

Adding Custom Rules

Add detection rules in src/rules/ruleEngine.ts:

{
    id: 'custom-rule',
    type: VulnerabilityType.CustomType,
    severity: vscode.DiagnosticSeverity.Warning,
    message: 'Description of the security issue',
    code: 'TC100',
    patterns: [/your-regex-pattern/g],
    languageIds: ['python', 'javascript', 'typescript'],
    documentation: 'How to fix this issue'
}

Roadmap

Planned for upcoming releases:

• [ ] Custom rule builder (YAML/JSON configuration)
• [ ] Quick-fix code actions for common issues
• [ ] CI/CD integration (GitHub Actions, GitLab CI)
• [ ] Compliance reporting (OWASP LLM Top 10, NIST AI RMF)
• [ ] Team policy enforcement
• [ ] Additional language support (Java, Go, C#, Rust)

Have a feature request? Open an issue.

Resources

• OWASP Top 10 for LLM Applications
• NIST AI Risk Management Framework
• OpenAI Security Best Practices

License

MIT License - see LICENSE for details.

Support

• Report a Bug
• Request a Feature
• Read the Docs

_{Built for developers building the next generation of AI applications}

⭐ Star on GitHub