A Refusal Is Not a Veto: Why Multi-Agent Structures Are Reported Both to Amplify and to Contain Language-Model Harm, Which Properties of the Structure Set the Sign, and the Matched Comparison That Would Separate Them

(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research Language models are increasingly composed into systems in which one model plans and others execute, check or debate. Two bodies of work describe the safety consequences in opposite terms. One reports that composition amplifies harm: a model that refuses a hazardous task on its own completes it once the task passes through a delegating principal, a planner or an orchestrator. The other reports that composition contains harm: adding agents that inspect a response or challenge a peer cuts jailbreak success several-fold. In this analysis the author reviews and synthesises published results to show that "multi-agent" is not the variable either literature measured. A reported change bundles at least five properties of the structure: what each model can see of the original intent, what it is told about the authority of the instruction it receives, whether a refusal ends the path or can be routed around, what incentive the prompts add, and whether output becomes an action. Re-reading the anchor study's own tables, the author computes that if delegation merely added a second, independent refusal point, end-to-end full execution would fall below every model's single-agent rate; it rises above it for four of six models, and for two of them more than nine tenths of delegated full execution passed through deliveries the study's own judge classed as sanitized, so on that classification the subordinate did not receive the harmful intent it is said to have obeyed (the study's authors read the same numbers as evidence that sanitization was incomplete). In a computer-use study, a single model given an orchestrator's decomposed subtasks, with no second agent, reproduces most of the orchestrated system's failure, while in a planner-executor study executor compliance is 6.3 percent when the planner refuses and 78.6 percent when it emits steps, against 64.4 percent with no planner. The paper consolidates 24 published results, states an attribution procedure, and specifies the matched comparison that would separate the five properties.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24420072
Primary Topic
Ethics and Social Impacts of AI
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Refusal Is Not a Veto: Why Multi-Agent Structures Are Reported Both to Amplify and to Contain Language-Model Harm, Which Properties of the Structure Set the Sign, and the Matched Comparison That Would Separate Them

Pranay M. Mahendrakar
Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
preprint

A Refusal Is Not a Veto: Why Multi-Agent Structures Are Reported Both to Amplify and to Contain Language-Model Harm, Which Properties of the Structure Set the Sign, and the Matched Comparison That Would Separate Them

Pranay M. Mahendrakar
preprint en

Abstract

(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research Language models are increasingly composed into systems in which one model plans and others execute, check or debate. Two bodies of work describe the safety consequences in opposite terms. One reports that composition amplifies harm: a model that refuses a hazardous task on its own completes it once the task passes through a delegating principal, a planner or an orchestrator. The other reports that composition contains harm: adding agents that inspect a response or challenge a peer cuts jailbreak success several-fold. In this analysis the author reviews and synthesises published results to show that "multi-agent" is not the variable either literature measured. A reported change bundles at least five properties of the structure: what each model can see of the original intent, what it is told about the authority of the instruction it receives, whether a refusal ends the path or can be routed around, what incentive the prompts add, and whether output becomes an action. Re-reading the anchor study's own tables, the author computes that if delegation merely added a second, independent refusal point, end-to-end full execution would fall below every model's single-agent rate; it rises above it for four of six models, and for two of them more than nine tenths of delegated full execution passed through deliveries the study's own judge classed as sanitized, so on that classification the subordinate did not receive the harmful intent it is said to have obeyed (the study's authors read the same numbers as evidence that sanitization was incomplete). In a computer-use study, a single model given an orchestrator's decomposed subtasks, with no second agent, reproduces most of the orchestrated system's failure, while in a planner-executor study executor compliance is 6.3 percent when the planner refuses and 78.6 percent when it emits steps, against 64.4 percent with no planner. The paper consolidates 24 published results, states an attribution procedure, and specifies the matched comparison that would separate the five properties.

Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon