New research reveals limits of LLM defenses against identity-linked decomposition attacks

A new research paper titled "Decomposition Attacks Across Unlinkable Identities: Limits of Stateful Defenses for LLM Services" explores the vulnerabilities of large language model services to decomposition attacks. These attacks split harmful tasks into multiple permissible requests, which stateless defenses can approve individually. The paper argues that effective defense requires stateful monitoring capable of grouping requests, but this becomes challenging when attackers use unlinkable identities. Experiments on numerous tasks and benign requests indicate that current defense policies struggle to stop these attacks without exceeding resource budgets, suggesting a need for additional mechanisms like identity linkage or control over answer usage. AI

IMPACT Highlights critical vulnerabilities in current LLM safety mechanisms, potentially requiring new defense strategies.

RANK_REASON Research paper published on arXiv detailing new attack vectors against LLM defenses. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• CORE Recommender
• DagsHub
• Decomposition Attacks Across Unlinkable Identities: Limits of Stateful Defenses for LLM Services
• Gotit.pub
• Hugging Face
• Influence Flower
• ScienceCast

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →