Red Teaming on Agentic AI

Sangdon Park
Sangdon Park
GPT-5

It is the era of Agentic AI but Agentic AI still suffers from trustworthiness issues, including weakness on jailbreaking. We advance the status of Agentic AI by debugging its issues via red teaming. In particular, we consider the following question:

How to learn red agentic AI to debug target agentic AI continuously?

We exploit our knowledge and experience on Trustworthy Agentic AI to build red Agentic AI for blue Agentic AI.

On-going/Potential Projects

  • Red Agent Learning: Learn red Agents for blue LLMs via RL.
  • Red Agent Continual Learning: Learn continual red teaming Agents.

Do you have creative ideas in building red Agentic AI?

Keywords

  • Red Teaming
  • Agentic AI
  • Reinforcement Learning