Tool Guard

​

What it does

• Analyzes request content for jailbreak signals (PreRequest)
• Uses an external firewall to score content risk
• Compares the maximum score against a threshold
• Actions when threshold is exceeded:
  • Block: return Forbidden (HTTP 403)
  • Throttle: delay the request, then allow
  • Alert only: allow and log/telemetry
• Adds a response header X-Jailbreak-Detected set to true/false

​

Configuration Parameters

• Stage: PreRequest
• If no content or parsing fails, the request is allowed (with telemetry)
• Uses a provider-specific parser internally (e.g., OpenAI)

​

Prerequisites

{
  "name": "{{upstream_service_name}}",
  "algorithm": "round-robin",
  "targets": [
    {
      "provider": "openai",
      "provider_options": { "api": "responses" },
      "weight": 50,
      "priority": 1,
      "default_model": "gpt-4o-mini",
      "models": ["gpt-4", "gpt-4o-mini"],
      "stream": false,
      "credentials": { "api_key": "" }
    }
  ]
}

​

Example configuration

{
  "name": "tool_guard",
  "enabled": true,
  "stage": "pre_request",
  "priority": 1,
  "parallel": false,
  "settings": {
    "mode": "block",
    "credentials": {
      "base_url": "https://firewall.example.com",
      "token": "${FIREWALL_TOKEN}"
    },
    "mapping_field": "prompt",
    "threshold": 0.7
  }
}

​

Best practices

• Start with alert_only to tune the threshold using real traffic
• Use mapping_field to point to relevant content (e.g., nested prompt fields)
• Pair with Tool Budget Limiter for cost control and layered agent security
• Monitor telemetry and headers to track detections over time

​

Compatibility