- Home
- Search Results
- Page 1 of 1
Search for: All records
-
Total Resources1
- Resource Type
-
00100000000
- More
- Availability
-
10
- Author / Contributor
- Filter by Author / Creator
-
-
Haghtalab, Nika (1)
-
Steinhardt, Jacob (1)
-
Wei, Alexander Wei (1)
-
#Tyler Phillips, Kenneth E. (0)
-
#Willis, Ciara (0)
-
& Abreu-Ramos, E. D. (0)
-
& Abramson, C. I. (0)
-
& Abreu-Ramos, E. D. (0)
-
& Adams, S.G. (0)
-
& Ahmed, K. (0)
-
& Ahmed, Khadija. (0)
-
& Aina, D.K. Jr. (0)
-
& Akcil-Okan, O. (0)
-
& Akuom, D. (0)
-
& Aleven, V. (0)
-
& Andrews-Larson, C. (0)
-
& Archibald, J. (0)
-
& Arnett, N. (0)
-
& Arya, G. (0)
-
& Attari, S. Z. (0)
-
- Filter by Editor
-
-
& Spizer, S. M. (0)
-
& . Spizer, S. (0)
-
& Ahn, J. (0)
-
& Bateiha, S. (0)
-
& Bosch, N. (0)
-
& Brennan K. (0)
-
& Brennan, K. (0)
-
& Chen, B. (0)
-
& Chen, Bodong (0)
-
& Drown, S. (0)
-
& Ferretti, F. (0)
-
& Higgins, A. (0)
-
& J. Peters (0)
-
& Kali, Y. (0)
-
& Ruiz-Arias, P.M. (0)
-
& S. Spitzer (0)
-
& Sahin. I. (0)
-
& Spitzer, S. (0)
-
& Spitzer, S.M. (0)
-
(submitted - in Review for IEEE ICASSP-2024) (0)
-
-
Have feedback or suggestions for a way to improve these results?
!
Note: When clicking on a Digital Object Identifier (DOI) number, you will be taken to an external site maintained by the publisher.
Some full text articles may not yet be available without a charge during the embargo (administrative interval).
What is a DOI Number?
Some links on this page may take you to non-federal websites. Their policies may differ from this site.
-
Large language models trained for safety and harmlessness remain susceptible to adversarial misuse, as evidenced by the prevalence of "jailbreak" attacks on early releases of ChatGPT that elicit undesired behavior. Going beyond recognition of the issue, we investigate why such attacks succeed and how they can be created. We hypothesize two failure modes of safety training: competing objectives and mismatched generalization. Competing objectives arise when a model's capabilities and safety goals conflict, while mismatched generalization occurs when safety training fails to generalize to a domain for which capabilities exist. We use these failure modes to guide jailbreak design and then evaluate state-of-the-art models, including OpenAI's GPT-4 and Anthropic's Claude v1.3, against both existing and newly designed attacks. We find that vulnerabilities persist despite the extensive red-teaming and safety-training efforts behind these models. Notably, new attacks utilizing our failure modes succeed on every prompt in a collection of unsafe requests from the models' red-teaming evaluation sets and outperform existing ad hoc jailbreaks. Our analysis emphasizes the need for safety-capability parity -- that safety mechanisms should be as sophisticated as the underlying model -- and argues against the idea that scaling alone can resolve these safety failure modes.more » « less