AI Red Teaming Techniques — Agentic Design

Overview#

9 Categories · 108 Techniques · 432 Attack Vectors · 12 Avg per Category

Attack Categories#

1. Prompt Injection (8 techniques) — Manipulate AI responses through malicious prompts
2. Jailbreaking (7 techniques) — Bypass AI safety mechanisms and content policies
3. Adversarial Attacks (2 techniques) — Create inputs designed to fool AI models
4. Vulnerability Assessment (4 techniques) — CVE analysis and security testing
5. Supply Chain Attacks (4 techniques) — Model and data supply chain security
6. Model Theft & IP Protection (5 techniques) — Model extraction and IP protection
7. Agentic AI Attacks (50 techniques) — Multi-agent security and autonomous exploitation
8. Memory & Context Attacks (13 techniques) — Memory poisoning, RAG exploitation
9. Multimodal Attacks (10 techniques) — Cross-modal exploitation

Complexity Distribution#

• Low: 4 techniques
• Medium: 42 techniques
• High: 62 techniques

Ethical Guidelines#

• Only test systems you own or have explicit permission to test
• Report vulnerabilities through proper channels
• Follow coordinated vulnerability disclosure
• Respect system availability and user privacy

Defensive Focus#

• Use techniques to improve system security
• Document and implement appropriate defenses
• Share knowledge to strengthen the AI security community
• Prioritize building safer, more robust AI systems