Cognitive Defense Agent

What It Does

The Cognitive Defense Agent uses a fine-tuned NLP model to detect prompt injection patterns, jailbreak attempts, role-playing attacks, and other adversarial text constructs embedded in data payloads. Critical for pipelines that feed into AI/LLM systems.

Capabilities

• Prompt injection detection
• Jailbreak attempt identification
• Role-playing attack detection
• Adversarial text classification
• Multi-language support
• Confidence scoring

Example

const result = await mcp.call('cognitive_defense_scan', {
  payload: normalizedPayload,
  stream_uuid: streamUUID,
  scan_depth: 'standard',
});
console.log(result.adversarial_detected); // false
console.log(result.confidence);           // 0.02

Configuration

agent: cognitive-defense-agent
version: "1.0"
model: redqueen-cognitive-defense-v2
scan_fields: all_text
detection:
  prompt_injection: true
  jailbreak: true
threshold:
  flag_confidence: 0.75
  reject_confidence: 0.95