ai-chat-protection

Validate every user message server-side with validatePromptServer before sending it to the LLM. It checks 49 injection patterns covering system prompt overrides, jailbreaks, and encoding attacks, returning a sanitized prompt only when validation succeeds.

Import validatePromptServer from lib/prompt-validation, call it with the message and user context inside your route handler, and return HTTP 400 on failure. Wrap the handler with withRateLimit and withCsrf middleware for layered protection.

validatePrompt performs basic validation of length, injection patterns, special characters, and repetition for client or simple server use. validatePromptServer adds rate limiting up to 100 requests per session and logs suspicious prompts containing words like admin, password, or token.

Yes, the validation detects hex, unicode, and URL-encoded payloads such as \x49\x67\x6e\x6f\x72\x65. It also rejects prompts where special characters exceed 30 percent of total length, which catches most obfuscated payloads.

Client-side validation only improves UX and can be bypassed by direct API calls. Attackers can send malicious payloads with curl, so the server must re-validate every prompt before it reaches the LLM.

Use policyPromptSchema, which blocks data access patterns like "show me all users" and "SELECT * FROM". Additionally, scope the AI context to only the authenticated user's own data fetched by their userId.