Reproduce “No Attacker Needed: Unintentional Cross-User Contamination in Shared-State LLM Agents” 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

Tiankai Yang, Jiate Li, Yi Nian, Shen Dong, Ruiyao Xu, Ryan Rossi, Kaize Ding, Yue Zhao

Shared-state LLM agents exhibit unintentional cross-user contamination (UCC), where benign interactions from one user lead to incorrect or silently erroneous outcomes for others, a previously understudied robustness problem. The research found contamination rates between 57% and 71% across different agent architectures, with text-level sanitization reducing this risk but proving insufficient when conventions are embedded in executable code.

Key insights

  • Unintentional cross-user contamination is a prevalent failure mode in shared-state LLM agents, occurring at high rates even from non-malicious inputs.
  • The effectiveness of text-level defenses like SSI varies significantly with the shared-state mechanism, performing well in conversational contexts but less so for states containing executable code.
  • Conventions embedded directly within executable artifacts (e.g., solution code) are resistant to text-level sanitization, requiring more sophisticated artifact-level defenses to prevent persistent contamination.

Summary by alphaXiv.

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