The Consensus Trap: Rescuing Multi-Agent LLMs from Adversarial Majorities via Token-Level Collaboration
Factual evidence
What the source reports
Research reveals multi-agent LLM systems using majority voting are vulnerable to adversarial prompt injections when corrupted agents outnumber benign ones.
Inspect the evidence
- Inclusion basis
- Enterprise AI
- Publisher and source type
- arXiv cs.CL — Computation and Language · RESEARCH
- Published by source
- 9 November 2026
- Collected by OneBench
- 21 Apr 2026, 07:28 UK
Stored source excerpt
arXiv:2604.17139v1 Announce Type: new Abstract: Multi-agent large language model (LLM) architectures increasingly rely on response-level aggregation, such as Majority Voting (MAJ), to raise reasoning ceilings.…
Short excerpt from the collected text, not the full source. Use the source link to read it in context.
The factual summary is a OneBench synthesis, not a quotation or independent verification. Collection time is not publication time. Open the source for its full context; related reporting can share the same underlying announcement.
OneBench interpretation
Institutional assessment
So what
This research identifies a critical vulnerability in multi-agent LLM architectures, which banks increasingly consider for complex reasoning tasks, directly impacting their security and reliability assessments.
Do what
This necessitates a re-evaluation of security protocols for any multi-agent LLM system currently in pilot or planning stages, particularly those relying on response-level aggregation.