Hidden AI Agent Communication Opens New Attack Vectors
Research shows that latent communication between AI agents can bypass safety measures, raising harmful compliance from 27.9 to 76.9 without modifying the models.

👁 118 reads
Research published on arXiv reveals a critical vulnerability in multi-agent AI systems: when models communicate through internal representations instead of text, their safety guardrails can be compromised without touching the models themselves.
Muhammad Huzaifa, Sina Mavali, and Thorsten Eisenhofer published their paper “Safety of Latent Communication in Multi-Agent Systems” on arXiv on September 30, 2026. They describe a problem that emerges when multiple AI models communicate not through text, but directly through internal representations - so-called latent communication.
What is latent communication between AI agents
Typically, AI agents exchange information through text, which generates costs: tokens are consumed, computation time increases, and latency grows. Latent communication bypasses this stage - one model passes its internal states directly into the input space of the receiver. To make this work, lightweight, trainable “links” are introduced between agents that map the sender’s representations into a format the receiver can understand.
The problem: even benign link training breaks safety
The authors show that simply training these links in a benign way - without any malicious intent - increases what’s called harmful compliance, or the system’s tendency to fulfill harmful requests. Importantly, the base models themselves remain unchanged and still have built-in safety alignment. The problem lies in the link, not in the agent.
An attacker can amplify this effect in two ways described in the paper: first, by optimizing links on pairs of harmful questions and harmful answers; second, by poisoning seemingly benign training data.
Attack through reinforcement learning
The authors developed a separate type of attack based on reinforcement learning. Its distinguishing feature is that it doesn’t require having target harmful responses - it rewards harmful compliance itself while maintaining utility on normal tasks.
The results are concrete: across three communication topologies and four safety benchmarks, this attack raised the average harmful compliance rate from 27.9 (with benignly trained links) to 76.9. Compared to direct supervised optimization, the RL attack also achieved higher average accuracy on two utility benchmarks - meaning the compromised system simultaneously performs normal tasks better.
Possible fix without modifying agents
The authors also tested whether the same mechanism can be used in the opposite direction. It turns out that adjusting rewards toward safe behavior allows compromised links to be repaired. According to the description in the abstract, this approach significantly reduces harmful compliance across all evaluated attacks - and importantly, does so without updating the agents themselves.
Main conclusion from the authors
The authors write explicitly: “safety alignment requires considering the multi-agent system as a whole” - securing a model in isolation is not enough if that model will operate as part of a larger multi-agent system. Communication links between agents become a separate attack surface that has not previously been considered in safety assessments.
The code for the paper is publicly available at the address indicated by the authors in the abstract. The article is classified in the categories Artificial Intelligence, Machine Learning, and Multiagent Systems.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →What this means
This research signals to us that multi-agent systems are entering a phase where the safety of a single model ceases to be a sufficient measure. The number 76.9 as a harmful compliance rate versus 27.9 at the starting point is a difference that cannot be dismissed. We find it particularly concerning that the RL attack doesn’t need ready-made harmful responses - this lowers the barrier to entry for potential abuse. At the same time, the fact that the same mechanisms can be used to repair links without touching the base models suggests practical defense paths - provided someone actually starts implementing them in production system design.
Source: arXiv cs.AI: Safety of Latent Communication in Multi-Agent Systems (document from 2026-10-02). This review was based on the full text of the document, not reports from other media.
More on this from QUANTSLANT
magazine·Why AI Open Source Is Strategy, Not Altruism: AstaBriefAllen AI released AstaBrief as open source. It's not charity - it's a calculated move in the battle for influence over AI infrastructure.

