<ul><li>LLM Flowbreaking is identified as the third attack type against LLMs after jailbreaking and prompt injection.</li><li>The 'Second Thoughts' attack involves the LLM retracting and replacing offensive content or displaying an error message.</li><li>By pressing the Stop button while the LLM is generating an answer, users can bypass guardrails and receive a violating response.</li><li>The attack focuses on exploiting the application architecture components surrounding the LLM, rather than the model itself.</li></ul>

Race Condition Attacks against LLMs

Discover more