Skip to Content
PatternsAI Safety

Patterns

/

AI Safety

Input validation, prompt injection detection, and rate limiting — ported from mukoko-weather’s production guards. Every Mukoko app integrating Claude or Shamwari AI must use these utilities.

Install from registry

npx shadcn@latest add https://registry.mukoko.com/api/v1/ui/ai-safety

Live demonstration

Test slug validation, prompt injection detection, and rate limiting interactively.

Slug Validation

Prompt Injection Detection

Rate Limiter

Sliding window rate limiter: 5 requests per 10 seconds. Click rapidly to hit the limit.

Input Validation

API route validation
import {
  validateSlug,
  validatePromptLength,
  sanitizeUserInput,
  validateArrayLength,
  validateAllowlist,
  truncateMessage,
} from "@/lib/ai-safety"
 
// Slug validation (URL-safe identifiers)
const { valid, sanitized } = validateSlug(params.slug)
if (!valid) {
  return Response.json({ error: "Invalid slug" }, { status: 400 })
}
 
// Message length (2000 char cap from mukoko-weather)
if (!validatePromptLength(body.message, 2000)) {
  return Response.json({ error: "Message too long" }, { status: 400 })
}
 
// Sanitize user input (remove control chars)
const cleanInput = sanitizeUserInput(body.prompt)
 
// Array limits (20 items max from mukoko-weather)
if (!validateArrayLength(body.activities, 20)) {
  return Response.json({ error: "Too many activities" }, { status: 400 })
}
 
// Tag allowlisting
const knownTags = ["farming", "mining", "travel", "tourism", "sports"]
const { valid: tagsOk, invalid, filtered } = validateAllowlist(body.tags, knownTags)
if (!tagsOk) {
  log.warn("Unknown tags filtered", { data: { invalid } })
}
 
// Truncate chat history (2000 chars per message)
const history = messages.map(m => ({
  ...m,
  content: truncateMessage(m.content, 2000),
})).slice(-10)  // max 10 messages

Prompt Injection Detection

Heuristic detection of common prompt injection patterns. Use as one layer in defense-in-depth alongside length limits, rate limiting, and system prompt guardrails. Not a guarantee — determined attackers will find bypasses.

prompt safety check
import { detectPromptInjection } from "@/lib/ai-safety"
 
// Check user input before sending to Claude
const { safe, flags } = detectPromptInjection(userMessage)
 
if (!safe) {
  logger.warn("Prompt injection attempt", {
    data: { flags, inputLength: userMessage.length },
  })
  return Response.json(
    { error: "Your message was flagged for safety review" },
    { status: 400 }
  )
}
 
// Detection categories:
// - system_override: "ignore previous instructions", "forget everything"
// - role_confusion: "pretend to be", "you are now a"
// - data_exfiltration: "reveal your system prompt", "repeat the text above"
// - delimiter_injection: "```system", "<system>" tags

Rate Limiting

In-memory sliding window rate limiter. Production limits from mukoko-weather:

EndpointLimitWindow
/api/py/chat20 req1 hour
/api/py/ai/followup30 req1 hour
/api/py/explore/search15 req1 hour
/api/py/history/analyze10 req1 hour
/api/py/locations/add5 req1 hour
/api/py/reports (submit)5 req1 hour
rate limiting in API routes
import { RateLimiter } from "@/lib/ai-safety"
 
// Create limiters per endpoint
const chatLimiter = new RateLimiter({ maxRequests: 20, windowMs: 3600_000 })
const searchLimiter = new RateLimiter({ maxRequests: 15, windowMs: 3600_000 })
 
// In API route handler
export async function POST(request: Request) {
  const ip = request.headers.get("x-forwarded-for") ?? "unknown"
  const { allowed, remaining, resetMs } = chatLimiter.consume(ip)
 
  if (!allowed) {
    return Response.json(
      { error: "Rate limit exceeded" },
      {
        status: 429,
        headers: {
          "Retry-After": String(Math.ceil(resetMs / 1000)),
          "X-RateLimit-Remaining": "0",
        },
      }
    )
  }
 
  // Process request...
}

AI integration checklist

  • Validate all slugs with validateSlug() before database queries
  • Cap message length to 2000 chars (validatePromptLength)
  • Sanitize user input (sanitizeUserInput) — remove control characters
  • Limit arrays to 20 items (validateArrayLength)
  • Validate tags against allowlist (validateAllowlist)
  • Check for prompt injection before sending to Claude (detectPromptInjection)
  • Rate limit all AI endpoints (RateLimiter)
  • Truncate chat history to 10 messages, 2000 chars each
  • Use system prompt guardrails — never let user override the system prompt
Last updated on
ResilienceChaos Testing