Only as Safe as the Weakest Agent: A Break That Appears Only When the System Is Multi-Agent
4.40T1 sourcearXiv cs.MA
arXiv cs.MAoriginal source ↗
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2611.17520.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryThe paper shows that multi-agent systems can fail even when their lead model resists attacks in isolation. By planting forged records in files read by weaker sub-agents, the authors induce harmful revisions and error propagation across Claude Code, Codex, a reference harness, and simulations of 18 models.
Why it mattersUseful for agent designers setting trust boundaries and selecting failure cases to test before deploying delegated sub-agents.
Cited by
No citations on record.