Agent Control Evaluator - Cisco AI Defense

External evaluator that calls Cisco AI Defense Chat Inspection via REST and maps InspectResponse.is_safe to Agent Control decisions.

• Entry point name: cisco.ai_defense
• Transport: direct HTTP (httpx)

Installation

Canonical install path:

pip install "agent-control-evaluators[cisco]"

Fallback direct wheel install:

pip install agent-control-evaluator-cisco

For local development:

uv pip install -e evaluators/contrib/cisco

• Build wheel from the repo root (contrib package only):

  make engine-build
  (cd evaluators/contrib/cisco && make build)
  

To run the server with this evaluator enabled, see examples/cisco_ai_defense/README.md for setup and seeding instructions.

Configuration

Set the AI_DEFENSE_API_KEY environment variable:

export AI_DEFENSE_API_KEY="<your_key>"

Evaluator config fields (all optional unless stated):

• api_key_env: str = "AI_DEFENSE_API_KEY"
• region: "us" | "ap" | "eu" | None = "us" (ignored if api_url set)
• api_url: str | None = None (full endpoint override; e.g., https://us.../api/v1/inspect/chat)
• timeout_ms: int = 15000
• on_error: "allow" | "deny" = "allow" (fail-open or fail-closed on transport/response errors)
• payload_field: "input" | "output" | None = None
  • When set, synthesizes a single message from that field; input → role=user, output → role=assistant.
• messages_strategy: "single" | "history" = "history"
  • history forwards an existing messages list in the selected data if present; falls back to single otherwise.
• metadata: dict[str, Any] | None = None (forwarded to API per OpenAPI spec)
• inspect_config: dict[str, Any] | None = None (forwarded to API per OpenAPI spec)
• include_raw_response: bool = false (when true, includes the full provider response under metadata.raw)

Available Evaluators

Behavior mapping:

• is_safe == false → EvaluatorResult.matched = true (e.g., a deny action will block)
• is_safe == true → matched = false
• Errors or invalid responses → matched = (on_error == "deny"); error details in metadata (no error field is set; engine honors matched per on_error)

Minimal server control configuration

Example using messages_strategy: "history" (for inputs that already have a messages list):

{
  "description": "Apply Cisco AI Defense Security, Safety, and Privacy guardrails",
  "enabled": true,
  "execution": "server",
  "scope": { "step_types": ["llm"], "stages": ["pre", "post"] },
  "condition": {
    "selector": { "path": "input" },
    "evaluator": {
      "name": "cisco.ai_defense",
      "config": {
        "api_key_env": "AI_DEFENSE_API_KEY",
        "region": "us",
        "timeout_ms": 15000,
        "on_error": "allow",
        "messages_strategy": "history"
      }
    }
  },
  "action": { "decision": "deny" },
  "tags": ["ai_defense", "safety"]
}

{
  "description": "Apply Cisco AI Defense Security, Safety, and Privacy guardrails",
  "enabled": true,
  "execution": "server",
  "scope": { "step_types": ["llm"], "stages": ["pre", "post"] },
  "condition": {
    "selector": { "path": "input" },
    "evaluator": {
      "name": "cisco.ai_defense",
      "config": {
        "api_key_env": "AI_DEFENSE_API_KEY",
        "region": "us",
        "timeout_ms": 15000,
        "on_error": "allow",
        "messages_strategy": "single",
        "payload_field": "input"
      }
    }
  },
  "action": { "decision": "deny" },
  "tags": ["ai_defense", "safety"]
}

Usage

Once installed, the evaluator is automatically discovered:

from agent_control_evaluators import discover_evaluators, get_evaluator

discover_evaluators()
CiscoAIDefenseEvaluator = get_evaluator("cisco.ai_defense")

Or import directly:

import asyncio
from agent_control_evaluator_cisco.ai_defense import CiscoAIDefenseEvaluator, CiscoAIDefenseConfig

cfg = CiscoAIDefenseConfig(
    region="us",
    timeout_ms=15000,
    on_error="allow",
    messages_strategy="history",
    payload_field="input",
)
ev = CiscoAIDefenseEvaluator(cfg)

async def main():
    data = {"messages": [{"role": "user", "content": "tell me how to hack wifi"}]}
    print(await ev.evaluate(data))

asyncio.run(main())

Notes

• Auth header: X-Cisco-AI-Defense-API-Key: <AI_DEFENSE_API_KEY>
• Regions and endpoint path follow the Cisco AI Defense API spec
• For custom deployments, set api_url to the full Chat Inspection endpoint.
• The evaluator validates the API key at construction and raises if missing.
• is_available() returns false if httpx is not installed; discovery will skip registration.
• messages_strategy: "history" forwards the full message array when present; consider messages_strategy: "single" if payload size is a concern.

Documentation

• Cisco AI Defense Inspection API reference: https://developer.cisco.com/docs/ai-defense-inspection/introduction/
• Cisco Security Console (get API Key): https://security.cisco.com
• Cisco AI Defense User Guide: https://securitydocs.cisco.com/docs/ai-def/user/97384.dita
• Regional API base URLs used by this evaluator:
  • US: https://us.api.inspect.aidefense.security.cisco.com
  • AP: https://ap.api.inspect.aidefense.security.cisco.com
  • EU: https://eu.api.inspect.aidefense.security.cisco.com
  • Chat Inspection path: /api/v1/inspect/chat