Junru Zhu’s SwarmSafe-Delta tests how messages affect agent safety

Junru Zhu’s SwarmSafe-Delta tests how messages affect agent safety

Junru Zhu introduces SwarmSafe-Delta, a protocol for testing whether messages change the committed actions of tool-using LLM agent teams. It holds upstream computation fixed and compares real message delivery with hidden delivery and format-matched placebo messages, then scores actions in benign and risky policy states under preassigned budgets. Across four models and constructed optional tasks, real delivery led to more risky-world violations than placebo or hidden delivery, while also improving benign task success. In a second regime constructed to require an upstream lookup route, real delivery improved safe completion relative to hidden delivery. These results show contrasting effects in the tested settings: communication can increase risky-world violations while also supporting task success, and suppressing it can sacrifice utility. Zhu stresses that the results quantify message-delivery contrasts under specified information dependencies, not naturally occurring failure rates. Exploratory analyses find that the effect depends strongly on whether an upstream message recommends a prerequisite check or direct execution; model differences also largely track those message compositions. The paper offers a controlled, execution-grounded measurement protocol, not a general claim that agent communication is harmful or protective.

Publication: 989a5a5c-5d5f-4d5e-84df-251dfd0e12c4

attribution — pass: The reading attributes the protocol, results and framing to Junru Zhu, the sole listed author of the arXiv paper. The results stay presented as the paper's reported findings.

Authors:Junru Zhu

summary — pass: The title, summary and question match the abstract. They cover the protocol design, the higher risky-world violations alongside improved benign success, the second lookup regime, the message-composition finding and the author's caveat about scope. Terms such as LLM agent, placebo delivery and preassigned budgets are dense. They are kept exactly, suit term-library tooltips, and the surrounding context is enough to follow the main idea. The question follows from the stated limits rather than leading the reader.

The contribution is a controlled, execution-grounded measurement protocol rather than a general claim that agent communication is harmful or protective.

date — pass: The original publication date of 2026-11-12 matches the v1 submission date on arXiv. The later feed date is correctly not used.

[Submitted on 12 Nov 2026]

support — pass: All claims and limitations rest on the abstract. The reading says communication can raise violations while supporting task success, and that suppressing it can sacrifice utility. Both points are stated in the abstract and confined to the tested settings. The disclaimer about naturally occurring failure rates is preserved. No unsupported benefits or forecasts are added.

These effects quantify message-delivery contrasts under specified information dependencies, not naturally occurring failure rates.

media — pass: The selected media is the verified direct link to the original arXiv abstract page. This is the most direct access to the source.