JailBound FOL-Guided Framework Evaluates LLM Jailbreaks
Original title: JailBound: FOL-Guided Jailbreak Evaluation for Exploring LLM Behavioral Boundaries
JailBound combines taxonomy-driven prompt generation with FOL-guided dual-branch search, boosting average ASR by 5.1 points across 46 LLMs.
LLM jailbreak benchmarks rely on manual prompts and lack systematic coverage. JailBound organizes threats by category and uses a meta-attack generator conditioned on taxonomy. FOL serves as sensitivity signal for high-value states, with High-Value and Safety-Sensitivity branches showing improved exploration.
- Taxonomy by risk category application domain attack type
- Fine-tuned meta-attack generator synthesizes prompts
- FOL-guided dual-branch search
- High-Value Branch improves ASR by 5.1 points
- Safety-Sensitivity Branch exhibits higher flip rates
- AI models
- LLMs
From the original article
1 min read at arxiv.orgAs large language models (LLMs) are increasingly deployed as general-purpose conversational assistants, they must handle diverse user requests while reliably refusing harmful ones. Jailbreak attacks craft adversarial instructions or conversational contexts to elicit unsafe responses despite these safeguards.
By Fazong Wu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.