Reproduce “Context-Fractured Decomposition Attacks on Tool-Using LLM Agents: Exploiting Artifact Provenance Gaps” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xiaofeng Lin, Yukai Yang, Daniel Guo, Sahil Arun Nale, Charles Fleming, Guang Cheng

Context-Fractured Decomposition (CFD) attacks, a novel class of multi-step jailbreaks, are introduced to exploit the "provenance gap" in tool-using LLM agents. These attacks leverage the lack of artifact history tracking across fragmented execution sessions to achieve harmful goals, demonstrating superior performance over existing jailbreaking methods with an average Attack Success Rate (ASR) of 66.03% on GPT-4.1.

Key insights

  • LLM agent systems are vulnerable to systemic attacks that exploit architectural weaknesses, such as fragmented execution and the absence of artifact provenance tracking, rather than solely linguistic or in-context prompt manipulations.
  • Decomposing a harmful objective into a series of locally benign, artifact-manipulating steps across multiple sessions effectively evades bounded-observability safety monitors, even highly capable ones.
  • The effectiveness of CFD attacks is directly correlated with increasing decomposition depth and width, indicating that more granular fracturing and delayed composition are crucial for bypassing enforcement boundaries.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.