Research map / Multi-agent systems and games / LLM agents
Cross-User Poisoning: User-Task Boundary Failures in Multi-User Collaborative Language Agents
Atharv Singh Patlan, Peiyao Sheng, S Ashwin Hebbar, Prateek Mittal, Pramod Viswanath
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Identifies a user-task boundary problem in multi-user collaborative language agents, where a shared agent must decide which user or task an instruction may govern, creating cross-user poisoning risks.
Problem
Shared language agents observe interleaved messages from multiple users, keep persistent context, and execute tool actions, so they must decide not only whether an instruction is safe but which user or task it is allowed to govern.
Approach
Frames cross-user poisoning as a boundary failure specific to multi-user shared agents, distinct from standard single-user agent safety concerns about whether an instruction is safe.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add task clustering (isolate inferred task context) (ablation) | GPT-4.1 shared agent · Slack, Workspace, Airline | ASR: 0.0 ASR in all three environments | Improved |
| add context augmentation (task-cluster IDs and summaries) (ablation) | GPT-4.1 shared agent · Slack, Workspace, Airline | ASR: 0.0 in Slack and Workspace, 5.7 in Airline | Improved |
| add selective summarization of other users' context (ablation) | GPT-4.1 shared agent · Slack, Workspace, Airline | ASR: 23-48% ASR remains; TSR lowered by 15-25 points | No clear effect |
| add task clustering (ablation) | GPT-4.1 shared agent · Slack, Workspace, Airline | TSR: Slack 81.0 -> 64.0; Workspace 72.1 -> 72.8; Airline 46.7 -> 41.1 | Hurt |
| add context augmentation (ablation) | GPT-4.1 shared agent · Slack, Workspace | TSR: Slack 81.0 -> 96.2; Workspace 72.1 -> 95.4 | Improved |
| add ProtectAI input filter against CUP | GPT-4.1 · MURMUR avg | ASR: 69.4 -> 43.3 | Mixed |
| add PromptGuard 2 input filter against CUP | GPT-4.1 · MURMUR | ASR: 42.7% ASR retained | Mixed |
| deliver attack objective via another user (CUP) instead of retrieved data (PI) | frontier models · Slack and Workspace | ASR: PI ASR below 5%; CUP nontrivial and often high | Improved |
| use instruction-detecting PI defenses on benign messages | DataSentinel, MELON, PromptGuard, GPT-5.4 judge · benign messages | false positive rate: 89.06, 100.00, 53.12, 35.53 percent flagged | Hurt |
| use TS-Guard action guard, aggressive threshold (risk >= 0.5) | GPT-4.1 · MURMUR | detection of CUP sessions: detects 85.4 while disrupting 27.93 of benign sessions | Mixed |
| replace main attacker template with minimal or natural-language template (ablation) | CUP on GPT-4.1 · Workspace, Slack, Airline | ASR: Slack 81.9 -> 64.4 (natural language) | Hurt |
| increase number of concurrent tasks | CUP attacks · MURMUR k up to 5 | compromised tasks: more tasks compromised as concurrency grows | Hurt |
Closest papers by meaning
- The Alignment Waltz: Jointly Training Agents to Collaborate for Safety (ICLR 2026 · similarity 0.87)
- DRIFT-BENCH: Diagnosing CoopeRative Breakdowns in LLM Agents under Input Faults via Multi-Turn Interaction (ICML 2026 · similarity 0.87)
- ChatInject: Abusing Chat Templates for Prompt Injection in LLM Agents (ICLR 2026 · similarity 0.86)
- Learning “Partner-Aware” Collaborators in Multi-Party Collaboration (NeurIPS 2025 · similarity 0.86)
- When Embedding-Based Defenses Fail: Rethinking Safety in LLM-Based Multi-Agent Systems (ICML 2026 · similarity 0.86)
- GUARDIAN: Safeguarding LLM Multi-Agent Collaborations with Temporal Graph Modeling (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
