← News

Agents & MCPSources: arXiv · 16 Nov 2026

Researchers propose ReMA attacks and SecMAS defense for multi-agent systems

An arXiv paper describes attacks that manipulate how AI agents answer subtasks, plus a secure aggregation rule with stated theoretical guarantees.

In plain words

Multi-agent systems split a task among several AI agents, then combine their answers. Malicious agents can skew the combined result. The paper introduces an attack method that refines malicious answers, and a defense that picks the answer most consistent with a large group of others.

What was announced

The authors frame response manipulation attacks as optimization problems, in both untargeted and targeted forms, and approximately solve them by iteratively refining malicious responses using feedback from a surrogate multi-agent system. This lets malicious agents shape the final solution a system produces from delegated subtasks.

As a countermeasure, the paper proposes SecMAS, an aggregation rule that uses semantic similarities among agent responses to select the response most consistent with a sufficiently large subset of the others. The authors report theoretical security guarantees against both attack types, and evaluations in which existing aggregation rules proved vulnerable to ReMA while SecMAS resisted all evaluated attacks, including adaptive variants, with utility comparable to existing rules.

What it does not show

  • The source is an abstract only; experimental setup, threat model details and full results are not available in the supplied passages.
  • Security guarantees and defense effectiveness are the authors' own claims and have not been independently confirmed.

For security and IT teams

Scope
The paper addresses multi-agent systems that delegate subtasks to agents and combine responses with an aggregation rule; it does not cover other architectures in the supplied evidence.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.