Agentic AI Top 10 Vulnerability (CSA/OWASP)

Top 10 threats and mitigation for AI Agents

Overview

This project documents the top 10 security risks specifically related to AI Agents, representing a comprehensive analysis of vulnerabilities unique to autonomous AI systems. The document provides detailed descriptions, examples, and mitigation strategies for each risk, helping organizations secure their AI agent deployments effectively.

Purpose

As AI agents become increasingly prevalent because GenAI models, understanding and mitigating their security risks becomes crucial. This guide aims to:

• Identify and explain the most critical security risks in AI agent systems
• Provide practical mitigation strategies for each identified risk
• Help organizations implement secure AI agent architectures
• Promote best practices in AI agent security

Project Structure

The documentation is organized into top ten main security risks, each covering a specific risk category:

1. AAI001: Agent Authorization and Control Hijacking
2. AAI011: Agent Untraceability
3. AAI002: Agent Critical Systems Interaction
4. AAI014: Agent Alignment Faking Vulnerability
5. AAI003: Agent Goal and Instruction Manipulation
6. AAI005: Agent Impact Chain and Blast Radius
7. AAI006: Agent Memory and Context Manipulation
8. AAI007: Agent Orchestration and Multi-Agent Exploitation
9. AAI009: Agent Supply Chain and Dependency Attacks
10. AAI012: Agent Checker out of the loop vulnerability

Future

1. AAI014: Agent Temporal Manipulation Time-based attacks
2. AAI015: Agent Inversion and Extraction Vulnerability
3. AAI016: Agent Covert Channel Explotation

Deprecated

1. AAI004: Agent Hallucination Exploitation (partial overlap with LLM Top 10, so letting it stay in the LLM Top 10)
2. AAI010: Agent Knowledge Base Poisoning (will be merged with AAI006 - Memory and Context manipulation)
3. AAI008: Agent Resource and Service Exhaustion (will be merged with AAI005 - Imapct Chain and Blast Radius)

Roadmap

• **Version 1.0 ** (Feb 2025)

...