Rahman et al.
Protected LLMs are far less vulnerable to multi-turn jailbreaks than benchmarks suggest
Rahman et al. show that a prompt-based safety layer cuts a strong baseline's attack success rate from 0.896 to 0.608, and propose ATTCK-R, a planning-driven framework that outperforms recent multi-turn jailbreaking methods on protected black-box LLMs.
Big Tech
Md Mushfiqur Rahman · Tobias Rohde · Rahul Tripathi · Khalil Elkhalil · Kevin Lybarger
George Mason University · Amazon
Research Digest··2 min read
1 percent relative drop.
Why this paper
From Amazon and George Mason University
In one line
ATTCK-R outperforms existing multi-turn jailbreaking methods against protected LLMs using planning and tactic curation.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§