A comprehensive, MITRE ATT&CK-inspired framework designed specifically to map out vulnerabilities and defensive strategies for autonomous LLM-based agents.
As AI agents gain more autonomy (browsing the web, executing code, interacting with other agents), they introduce entirely new attack surfaces. This matrix provides a structured approach to securing them.
The matrix is divided into 5 core tactical phases:
- Reconnaissance & Initial Access (e.g., Goal Hijacking, NHI Delegation)
- Execution & Tool Access (e.g., CoT Poisoning, Browser DOM Isolation)
- Persistence & Lateral Movement (e.g., Episodic Memory Subversion, Semantic Kill-Switches)
- Impact & Data Exfiltration (e.g., Context Exhaustion, Secret Masking)
- Detection, Response & Governance (e.g., Agent UEBA, Multi-Agent Reputation)
π View the Interactive Matrix
This is an evolving framework. If you are working on AI Red Teaming or Agentic Security, please open an Issue or submit a PR to add new attack vectors and mitigations!