/prompt-injection-detector

name: prompt-injection-detector
description: Prompt injection detection and prevention for secure LLM applications
allowed-tools:
  - Read
  - Write
  - Edit
  - Bash
  - Glob
  - Grep
graph:
  domains: [domain:software-engineering]
  specializations: [specialization:ai-agents-conversational]
  skillAreas: [skill-area:hallucination-mitigation-fact-checking, skill-area:safety-redteaming]
  roles: [role:ml-engineer, role:backend-engineer]
  workflows: [workflow:feature-development, workflow:ml-model-lifecycle]

Prompt Injection Detector Skill

Capabilities

• Detect prompt injection attempts
• Implement input sanitization
• Configure detection classifiers
• Design defense layers
• Implement canary token detection
• Create injection logging and alerting

Target Processes

• prompt-injection-defense
• tool-safety-validation

Implementation Details

Detection Methods

1. **Pattern Matching**: Known injection patterns 2. **ML Classifiers**: Trained injection detectors 3. **Canary Tokens**: Detect instruction override 4. **LLM-Based**: Use LLM to detect manipulation 5. **Perplexity Analysis**: Unusual input patterns

Defense Strategies

• Input preprocessing
• Prompt structure design
• Output validation
• Sandboxed execution
• Multi-layer defense

Configuration Options

• Detection threshold
• Pattern rules
• Classifier model
• Action policies
• Alerting settings

Best Practices

• Defense in depth
• Regular pattern updates
• Monitor false positives
• Test with red-team inputs

Dependencies

• rebuff (optional)
• transformers
• Custom classifiers