Skip to main content
The MeshAI proxy scans every outbound LLM request for prompt injection patterns. When a match is detected, the proxy returns a 403 response and logs an audit event — the malicious prompt never reaches the LLM provider.

​
How It Works

  1. Every request flowing through the proxy is scanned against 15+ prompt injection patterns.
  2. If a pattern matches, the proxy blocks the request with 403 Forbidden.
  3. An policy.violated audit event is logged with policy_type: prompt_injection.
  4. The original prompt is stored in the audit trail for forensic review.

​
Patterns Detected

The scanner catches a wide range of known injection techniques:
CategoryExamples
Role override”Ignore all previous instructions”, “You are now…”
System prompt extraction”Repeat your system prompt”, “What are your instructions?”
Jailbreak attempts”DAN mode”, “Developer mode enabled”
Delimiter attacks---END SYSTEM---, [INST] injection
Encoding evasionBase64-encoded instructions, Unicode tricks
Context manipulation”Forget everything above”, “New conversation”
Nested injectionInstructions hidden inside data payloads

​
Configure via Policy

Prompt injection detection is controlled through the standard policy system. Create a prompt_injection policy to enable it: 
curl -X POST https://api.meshai.dev/governance/policies \
  -H "Authorization: Bearer msh_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "block-prompt-injection",
    "policy_type": "prompt_injection",
    "config": {
      "enabled": true,
      "sensitivity": "high"
    },
    "scope": "global",
    "enabled": true
  }'

{
  "success": true,
  "data": {
    "id": "pol_pi_001",
    "name": "block-prompt-injection",
    "policy_type": "prompt_injection",
    "config": {
      "enabled": true,
      "sensitivity": "high"
    },
    "scope": "global",
    "enabled": true,
    "created_at": "2026-03-18T10:00:00Z"
  }
}

​
Sensitivity Levels

LevelBehavior
lowCatches obvious injection patterns (role override, jailbreak keywords)
mediumAdds delimiter attacks, encoding evasion, and context manipulation
highAll patterns including nested injection and heuristic detection

​
What the Agent Sees

When a prompt injection is detected, the agent receives: 
{
  "error": "Policy violation: prompt_injection — request blocked due to detected prompt injection pattern",
  "policy_id": "pol_pi_001",
  "status": 403
}

​
Audit Trail

Every blocked injection attempt is logged:
Event TypeDescription
policy.violatedPrompt injection detected (includes matched pattern category and request metadata)
Query injection-related audit events: 
curl "https://api.meshai.dev/governance/audit-trail?event_type=policy.violated&policy_type=prompt_injection" \
  -H "Authorization: Bearer msh_YOUR_API_KEY"

​
Enforcement Architecture

  • Scan path: Proxy scans all messages content before forwarding to the LLM provider.
  • Latency: Pattern matching adds < 2ms overhead per request.
  • Fail-open: If the scanner errors, the request proceeds (logged as a warning).
  • No data exfiltration: Blocked prompts are stored only in MeshAI audit trail, never forwarded.