Summary
- This document details several test cases for evaluating the behavior of large language models used in multi-agent systems.
- The first test case, inter-agent prompt injection, examines whether an injected prompt in one agent’s output leads to the execution of unintended instructions by the next agent.
- The second test case, hallucination propagation, investigates the spread of fabricated facts through different agents.
- The third test case, data leakage via shared memory, focuses on personal information or secrets inadvertently shared between agents through shared memory or persistent context.
- These test cases can help developers identify potential threats, misbehaviors, or failure points in a multi-agent system powered by LLMs.
It is important to note that the document should solely be used for educational purposes and that any form of unauthorized system exploitation is illegal and punishable by law.