Skip to main content

BlackRainbow

41 techniques across 8 categories for AI red teaming. Each technique includes framework mappings to OWASP LLM Top 10, MITRE ATLAS, and F.O.R.G.E..

Categories

CategoryTechniquesFocus
Reconnaissance8Mapping AI system attack surface: model identification, prompt extraction, API enumeration
Prompt Injection7Injecting instructions into LLM context: direct override, framing, smuggling, indirect
RAG Attacks5Targeting retrieval pipelines: knowledge poisoning, retrieval hijacking, exfiltration
Agent & MCP6Exploiting agent frameworks: ReAct injection, tool poisoning, confused deputy, rug pulls
Adversarial ML4Attacking models directly: evasion, data poisoning, membership inference, extraction
Evasion4Bypassing security controls: unicode, token boundaries, output filters, payload splitting
Infrastructure3Targeting AI infra: model server CVEs, vector database attacks, deserialization RCE
Tools & Frameworks4Open-source red teaming tools: promptfoo, PyRIT, DeepTeam, Garak

Framework Mappings

Every technique maps to one or more industry frameworks:

  • OWASP Top 10 for LLMs — Vulnerability taxonomy for LLM applications
  • OWASP ML Top 10 — Machine learning-specific security risks
  • MITRE ATLAS — AI threat matrix (extends ATT&CK for ML systems)
  • F.O.R.G.E. — AI-integrated security assurance techniques (forged.itsbroken.ai)
  • NVIDIA AI Kill Chain — AI system attack lifecycle methodology
  • Google SAIF — Secure AI Framework for organizational posture

How to Use This

For red teamers: Work through categories sequentially during an AI system engagement. Start with Reconnaissance to map the target, then select attack techniques based on what you find.

For blue teamers: Use the framework mappings to prioritize defenses. Each technique page includes the attack surface it targets, helping you build detection rules.

For builders: Review techniques relevant to your architecture. If you're building a RAG pipeline, review the RAG Attacks category. If you're deploying agents, review Agent & MCP.

Attack Surface Map


BlackRainbow — an itsbroken.ai project.