AgentSkillsCN

security-red-teaming

系统性地对人工智能系统进行安全漏洞与攻击向量检测。

SKILL.md
--- frontmatter
name: security-red-teaming
description: Systematically test AI systems for security vulnerabilities and attack vectors.

Security Red Teaming Skill

Use this skill when conducting comprehensive security testing of AI systems, especially conversational AIs.

Attack Vector Categories

  1. •Prompt Injection: Attempts to override system instructions
  2. •Jailbreaking: Trying to break character or role (DAN mode, developer mode)
  3. •Data Exposure: Attempts to extract system information, API keys, or internal data
  4. •Input Manipulation: Encoding tricks, long inputs, XSS attempts
  5. •Role Confusion: Identity confusion, role-playing attacks
  6. •Technical Probes: SQL injection, system architecture questions

Testing Methodology

  1. •

    Plan Attack Vectors

    • •Create comprehensive list of potential vulnerabilities
    • •Categorize by severity and likelihood
    • •Document expected vs actual behavior
  2. •

    Execute Systematic Tests

    • •Test each attack vector methodically
    • •Document exact input and complete response
    • •Note any security boundary breaches
  3. •

    Analyze Results

    • •Identify successful vs failed attacks
    • •Assess response consistency and effectiveness
    • •Measure time to respond and response quality
  4. •

    Strengthen Defenses

    • •Update system prompts with improved guardrails
    • •Add specific deflection strategies for common attacks
    • •Implement input validation and sanitization
  5. •

    Retest and Validate

    • •Re-run attack vectors after improvements
    • •Ensure legitimate functionality remains intact
    • •Document security improvements

Key Principles

  • •Comprehensive Coverage: Test all known attack vectors systematically
  • •Response Consistency: Ensure uniform, secure responses to all attack types
  • •User Experience: Maintain helpfulness for legitimate queries
  • •Documentation: Keep detailed logs of all security tests and responses

Common Patterns

  • •Deflection Language: Use varied, warm responses when rejecting attacks
  • •Information Boundaries: Never reveal system internals, API details, or training data
  • •Character Consistency: Maintain persona even under attack
  • •Graceful Degradation: Handle malformed inputs without crashing

Success Metrics

  • •100% Attack Resistance: All security tests should fail safely
  • •Response Quality: Maintain helpfulness for legitimate users
  • •Performance: No degradation in response time or quality
  • •Consistency: Uniform security posture across all interaction types</content> <parameter name="filePath">/home/koo/github/Luxuryshoppingwebsite/.github/skills/security-red-teaming/SKILL.md