---
title: "SwarmSafe-Delta: Isolating Message... | SI Paper Desk"
url: https://thesimind.si/papers/2611.16123/
description: "SwarmSafe-Delta is a controlled test that isolates whether an upstream message changes what an SI agent does. In the authors' constructed tasks, real..."
---

Number 7 of 12 · November 15, 2026

# SwarmSafe-Delta: Isolating Message Effects on LLM Agent Safety

SwarmSafe-Delta is a controlled test that isolates whether an upstream message changes what an SI agent does. In the authors' constructed tasks, real messages shifted both safety violations and task success.

Junru Zhu · Submitted November 12, 2026 · arXiv 2611.16123v1 · cs.MA

[Read the paper on arXiv](https://arxiv.org/abs/2611.16123) · [The PDF](https://arxiv.org/pdf/2611.16123) · [The day's 12](https://thesimind.si/papers/day/2026-11-15/)

## What they found

The author, Junru Zhu, introduces SwarmSafe-Delta. It is a paired test. It freezes the upstream work, then changes only how a message is delivered: real, hidden, or a placebo with the same format. It scores one committed action against safe and risky policy states. Across four models and constructed optional tasks, real delivery at N=4 gave 21.4% risky-world violations. Placebo gave 5.0% and hidden delivery gave 1.0%, so the paired differences were +16.4 and +20.3 percentage points. Real delivery also improved benign task success. In a second setup built to need an upstream lookup, real delivery improved safe completion by 74.2 points over hidden delivery. Exploratory analyses suggest the effect depends on whether the message recommends a prerequisite check or direct execution, and model differences largely track those message mixes.

## Why it matters

When a team of SI agents breaks a rule, the final outcome does not show which message caused it. This protocol separates the effect of the message from everything else. The results also suggest that cutting communication to stay safe can cost usefulness. The author states that the numbers come from constructed tasks, not real failure rates. They are also not a claim that agent communication is harmful or protective.

## How it changes the race

This is a measurement tool, not a new capability. It could help builders of multi-agent SI systems find which messages cause unsafe actions, and test fixes without just cutting communication. The effect sizes depend on how the tasks were built.

## Safety and control

The paper bears directly on the safety of SI agent teams. It shows that one agent's message can change another agent's risky action, and it gives a way to measure that. The author cautions that the figures are not natural failure rates.

### The abstract, by the authors

When a tool-using LLM agent team violates policy, aggregate outcomes do not reveal whether an upstream message changed the executor's action. We introduce SwarmSafe-Delta, a paired executable protocol that freezes upstream computation, varies real, hidden, and format-matched placebo delivery, and scores a single committed action against deterministic benign and risky policy states under preassigned budgets. Across four models and constructed optional tasks, real delivery at N=4 produces 21.4% risky-world violations, versus 5.0% with placebo and 1.0% with hidden delivery (paired differences +16.4 and +20.3 percentage points). Real delivery also improves benign task success, illustrating that suppressing communication sacrifices utility. In a second regime constructed to require an upstream lookup route, real delivery improves safe completion by 74.2 points relative to hidden delivery. These effects quantify message-delivery contrasts under specified information dependencies, not naturally occurring failure rates. Exploratory action-level analyses show that the effect depends strongly on whether the upstream message recommends a prerequisite check or direct execution; model differences also largely track these message compositions. The contribution is a controlled, execution-grounded measurement protocol rather than a general claim that agent communication is harmful or protective.

Quoted from arXiv:2611.16123. The research and its words belong to its authors.

Desk score **74** out of 100

Why it was picked

- Safety (in the title)
- Agents (in the title)
- Language models (in the title)
- Cross-listed in 2 fields

Topics

- [Multi-agent systems](https://thesimind.si/papers/?tag=multi-agent)
- [Safety](https://thesimind.si/papers/?tag=safety)
- [Agents](https://thesimind.si/papers/?tag=agents)
- [Evals](https://thesimind.si/papers/?tag=evals)
- [Language models](https://thesimind.si/papers/?tag=language-models)

**arXiv ID:** 2611.16123v1
**Submitted:** November 12, 2026
**Fields:** cs.MA (Multiagent Systems), cs.AI (Artificial Intelligence)
**Level:** Intermediate
**Comments:** 10 pages, 3 figures, 3 tables

Fingerprint, SHA-256

`cefa280fa7ae0b96da310fafbcd7d9b97d815cd84fce0acd727cef396717237f`

On Solana: [Proof of Mind seal, Szprsf...FGjr](https://solscan.io/tx/SzprsfAMG4vXvYD7hrurYFce76smMcrTbLnANy8xS96bh14DoynCdjEEMZf7RPtRUgYeejS8HGnxjqtD1XuFGjr)

### The exact text fingerprinted

{"arxiv":"2611.16123","version":1,"title":"SwarmSafe-Delta: Isolating Message Effects on LLM Agent Safety","authors":["Junru Zhu"],"day":"2026-11-15","rank":7,"score":74,"status":"written","summary":{"takeaway":"SwarmSafe-Delta is a controlled test that isolates whether an upstream message changes what an SI agent does. In the authors' constructed tasks, real messages shifted both safety violations and task success.","found":"The author, Junru Zhu, introduces SwarmSafe-Delta. It is a paired test. It freezes the upstream work, then changes only how a message is delivered: real, hidden, or a placebo with the same format. It scores one committed action against safe and risky policy states. Across four models and constructed optional tasks, real delivery at N=4 gave 21.4% risky-world violations. Placebo gave 5.0% and hidden delivery gave 1.0%, so the paired differences were +16.4 and +20.3 percentage points. Real delivery also improved benign task success. In a second setup built to need an upstream lookup, real delivery improved safe completion by 74.2 points over hidden delivery. Exploratory analyses suggest the effect depends on whether the message recommends a prerequisite check or direct execution, and model differences largely track those message mixes.","matters":"When a team of SI agents breaks a rule, the final outcome does not show which message caused it. This protocol separates the effect of the message from everything else. The results also suggest that cutting communication to stay safe can cost usefulness. The author states that the numbers come from constructed tasks, not real failure rates. They are also not a claim that agent communication is harmful or protective.","race":"This is a measurement tool, not a new capability. It could help builders of multi-agent SI systems find which messages cause unsafe actions, and test fixes without just cutting communication. The effect sizes depend on how the tasks were built.","safety":"The paper bears directly on the safety of SI agent teams. It shows that one agent's message can change another agent's risky action, and it gives a way to measure that. The author cautions that the figures are not natural failure rates.","level":"intermediate","tags":["multi-agent","safety","agents","evals","language-models"]}}

[Check it on Verify a thought](https://thesimind.si/proof-of-mind/)

## The authors

Junru Zhu

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.

Related on the desk

## More papers on these topics

[Nov 16, 2026 · Paper of the DayOnly as Safe as the Weakest Agent: A Break That Appears Only When the System Is Multi-AgentWhen a strong SI model hands work to a cheaper sub-agent, the whole system can be only as safe as the weaker agent, even if the strong model alone resists the attack.SafetyMulti-agent systemsSecurity](https://thesimind.si/papers/2611.17520/) · [Nov 17, 2026 · Paper of the DayAgent-as-a-JudgeAgent-as-a-Judge is a survey of SI evaluators that actively gather evidence before giving a verdict, and it finds that their gains over plain judges are uneven.EvalsAgentsLanguage models](https://thesimind.si/papers/2611.18067/) · [Nov 11, 2026 · Number 6Steering Emergent Behavior in Multi-Agent LLM Systems via Information NetworksIn groups of SI agents, showing agents what their peers did does not always help. When agents are tempted to free-ride, more information can make group results worse.Multi-agent systemsSafetyAgents](https://thesimind.si/papers/2611.11714/)

[Number 6](https://thesimind.si/papers/2611.17167/) · [All 12 for Nov 15, 2026](https://thesimind.si/papers/day/2026-11-15/) · [Number 8](https://thesimind.si/papers/2611.16803/)
