no-training-data-exposure

Prevents user data from being sent to LLM training endpoints.

📊 Rule DetailsCopy Anchor Link

🔍 What This Rule DetectsCopy Anchor Link

This rule identifies code patterns where user data might be sent to LLM training endpoints or when training data collection is enabled.

❌ Incorrect CodeCopy Anchor Link

// Training enabled
const config = {
  training: true,
};

// Allow training flag
const options = {
  allowTraining: true,
};

// Training endpoint
fetch('https://api.openai.com/v1/fine-tune');

✅ Correct CodeCopy Anchor Link

// Training disabled
const config = {
  training: false,
};

// No training endpoint
await generateText({
  model: openai('gpt-4'),
  prompt: userInput,
});

⚙️ OptionsCopy Anchor Link

🛡️ Why This MattersCopy Anchor Link

Exposing user data to training can:

• Privacy violations - User data used without consent
• Data poisoning - Malicious data taints model
• Compliance violations - GDPR, CCPA violations
• IP leakage - Proprietary information exposed

🔄 Changes in the 2026-10-10 FP/FN auditCopy Anchor Link

• URLs are matched by whole path segment (/v1/fine_tuning/jobs, /api/train/model), not substring: /trainers and /training-schedule are not reported.
• feedback, improve and learn were removed from the defaults: a thumbs-up/down /api/feedback endpoint and showFeedback: true are product features, not training opt-ins. Flag names are matched by whole word.
• Now off in recommended: the rule has no tie to a provider setting.
• Options: the documented defaultOptions now apply. Earlier versions read context.options and silently ran a shorter hard-coded list instead.

Known False NegativesCopy Anchor Link

The following patterns are not detected due to static analysis limitations:

Environment-Based Training FlagsCopy Anchor Link

Why: Environment variables are not resolved.

// ❌ NOT DETECTED - Training from env
const options = { training: process.env.ENABLE_TRAINING };

Mitigation: Hardcode training: false. Never use env for training flags.

Training Endpoints in ConfigCopy Anchor Link

Why: Endpoints from config files are not visible.

// ❌ NOT DETECTED - Endpoint from config
fetch(config.apiEndpoint); // May be fine-tune endpoint

Mitigation: Review API configurations for training endpoints.

Implicit Training via SDK OptionsCopy Anchor Link

Why: Hidden SDK options enabling training may not be detected.

// ❌ NOT DETECTED - SDK defaults to training
const client = new AIClient(); // training: true by default

Mitigation: Explicitly set training: false in all SDK configs.

📚 ReferencesCopy Anchor Link

• OWASP LLM03: Training Data Poisoning
• CWE-359: Privacy Violation

Use this rule

npm install -D oxlint-plugin-vercel-ai-security

{  "jsPlugins": ["oxlint-plugin-vercel-ai-security"],  "rules": {    "vercel-ai-security/no-training-data-exposure": "error"  }}

import { defineConfig } from 'vite-plus';export default defineConfig({  lint: {    jsPlugins: ['oxlint-plugin-vercel-ai-security'],    rules: {      'vercel-ai-security/no-training-data-exposure': 'error',    },  },});