[WIP] Jailbreak Paradox: The Achilles' Heel of LLMs
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2024
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866910496814268416 |
|---|---|
| author | Rao, Abhinav Choudhury, Monojit Aditya, Somak |
| author_facet | Rao, Abhinav Choudhury, Monojit Aditya, Somak |
| contents | We introduce two paradoxes concerning jailbreak of foundation models: First, it is impossible to construct a perfect jailbreak classifier, and second, a weaker model cannot consistently detect whether a stronger (in a pareto-dominant sense) model is jailbroken or not. We provide formal proofs for these paradoxes and a short case study on Llama and GPT4-o to demonstrate this. We discuss broader theoretical and practical repercussions of these results. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2406_12702 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | [WIP] Jailbreak Paradox: The Achilles' Heel of LLMs Rao, Abhinav Choudhury, Monojit Aditya, Somak Computation and Language We introduce two paradoxes concerning jailbreak of foundation models: First, it is impossible to construct a perfect jailbreak classifier, and second, a weaker model cannot consistently detect whether a stronger (in a pareto-dominant sense) model is jailbroken or not. We provide formal proofs for these paradoxes and a short case study on Llama and GPT4-o to demonstrate this. We discuss broader theoretical and practical repercussions of these results. |
| title | [WIP] Jailbreak Paradox: The Achilles' Heel of LLMs |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2406.12702 |