Source: AgentDojo
Debenedetti et al. (2024): a dynamic benchmark of 97 agent tasks and 629 prompt-injection test cases in simulated email, banking and travel environments.
Contents
- What: benchmark paper, arXiv:2406.13352
- Authors: Edoardo Debenedetti, Jie Zhang, Mislav Balunović, Luca Beurer-Kellner, Marc Fischer, Florian Tramèr
- Date: 2024-06-19 (v3 2024-11-24)
- Original: arXiv:2406.13352
Summary
An extensible environment for measuring both utility and injection robustness of tool-using agents. Tools return data that may contain attacker instructions; the benchmark scores whether the user task succeeds and whether the attacker’s goal is achieved.
Key claims
- 97 realistic tasks and 629 security test cases in simulated email, e-banking and travel-booking settings.
- Even without attacks, state-of-the-art LLMs fail many tasks.
- Existing attacks break some but not all security properties; new attacks, defences and adaptive attacks can be plugged in.
- Later used as the evaluation suite for CaMeL (see CaMeL).