Top 10 threats and mitigation for AI Agents
A documented top-ten list of security risks specific to autonomous AI agent systems, such as authorization hijacking, memory manipulation and multi-agent orchestration exploitation, with mitigation strategies for each.
This project documents ten security risks specific to AI agents, forming the core of an OWASP and Cloud Security Alliance red-teaming reference distinct from the general LLM Top 10. Each risk, for example agent authorization and control hijacking, agent untraceability, critical-systems interaction, alignment-faking, goal and instruction manipulation, and orchestration or supply-chain attacks, is described with detailed examples and mitigation strategies aimed at securing real agent deployments.
The project tracks its own evolution: some risks are marked deprecated and folded into related categories, such as hallucination exploitation moving into the general LLM Top 10 and knowledge-base poisoning merging into memory manipulation, and future risks such as temporal manipulation and covert-channel exploitation are already listed on the roadmap.