AgentDojo presents a dynamic environment for assessing the security of AI agents against prompt injection attacks. It offers a range of tasks and security scenarios to evaluate the resilience of LLM agents. The framework aims to foster research on designing reliable and secure AI agents for common tasks.

• Evaluation framework for LLM agents’ robustness
• Realistic tasks and security test cases
• Challenges for attacks and defenses
• Potential for designing secure AI agents

The framework contributes to enhancing the security posture of AI agents and promotes the development of more resilient systems.