cs.AISep 18, 2025

Sentinel Agents for Secure and Trustworthy Agentic AI in Multi-Agent Systems

Authors: Diego Gosmar, Deborah A. Dahl

Organizations: Head of AI Tesisquare Voiceinteroperability.ai Initiative Member Linux Foundation AI & Data Torino, TO 10100, Italy · Principal Conversational Technologies Voiceinteroperability.ai Initiative Member Linux Foundation AI & Data Plymouth Meeting, Pennsylvania, USA

Abstract

Multi-agent systems (MAS) built on large language models (LLMs) increasingly share a common conversational space, in which a single adversarial or erroneous message can propagate from one agent to many. This paper proposes Sentinel Agents, a dedicated security layer that monitors inter-agent communication with rule-based checks, behavioral analysis and LLM-based semantic inspection, and reports to a Coordinator Agent that enforces policy, for example by quarantining misbehaving agents. We describe the architecture, its deployment patterns and monitoring layers, and we evaluate a continuous-listener prototype on a balanced corpus of 1,000 synthetic messages (500 attacks across five families and 500 benign messages) with two local LLMs. LLM inspection with Gemma 4 12B reached F1 0.86 at a false-positive rate of 0.11, against 0.38 for rules alone. Combining components by logical OR did not improve on it, while a cascade in which only high-precision rules act on their own matched it. Calibrating the thresholds of the more conservative Qwen3 8B on a development split raised its held-out F1 from 0.56 to 0.80, and on multi-turn conversations the behavioral check detected patterns the rules missed but could not tell legitimate repetition from probing. The value of combining detectors therefore depends on the model, on its calibration and on the aggregation rule, and the results yield practical guidance for deploying Sentinel Agents.

Figures & tables

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Oct 20, 2026

MAS-Guard: Unified Guardrails from Individual Agents to Multi-Agent LLM Systems

Agentic AI extends large language models (LLMs) from generating responses to planning and executing complex workflows, substantially expanding their utility while exposing applications to unsafe actions, information leakage, and adversarial manipulation. Multi-agent systems further expand these capabilities through specialization and collaboration, but distributed authority, heterogeneous roles, and opaque or potentially compromised participants make safeguarding even more challenging. Existing defenses address important individual threats but provide fragmented coverage across local execution, system-level interactions, and enforcement mechanisms. We introduce MAS-Guard, a unified framework that closes this gap for safeguarding single- and multi-agent AI applications containing both instrumentable and opaque agents. Within instrumentable agents, the proposed local Agent Guardrail (A-Guard) layer applies task-specific experts to assess agentic events. At the system level, developer-owned Agent Charters and a Control Plane combine deterministic structural checks of routes, agent capabilities, and message formats with semantic assessments of roles and message content. These complementary layers enable enforcement through a shared set of interventions: allow, reflection, confirmation, and block. Evaluated across six benchmarks covering diverse single- and multi-agent scenarios with three agent LLMs from different providers, MAS-Guard reduces the overall attack success rate by 71.5-86.3% relative to unguarded agents, while meeting the utility reference criterion on benign tasks.
Sep 1, 2026

SoK: When Safe Agents Fail Together: The Security of Multi Agent LLM Systems

Safe agents can fail together. Multi-agent LLM systems (MAS) move information, state, decisions, and authority across principal boundaries, creating failures that local checks may miss. Without an execution-level view, a multi-agent setting can easily be mistaken for evidence of a genuinely multi-agent security effect. We thus systematize MAS security through an execution-centered analysis of 197 works, covering six interaction interfaces, four adversary positions, seven system-level risks, and eight recurring attack paths. We introduce an A-I-R framework that organizes attacks by adversary position, interaction interface, and resulting system-level risk, unifying otherwise fragmented attack mechanisms across MAS. We organize defenses through a five-part contract covering path target, observation, intervention, trust boundary, and recovery, and identify path closure and recovery as key challenges. We audit 44 evaluation and benchmark works and identify open challenges in isolating interaction effects, designing comparable and diagnostic metrics, supporting reuse across MAS designs, and evaluating open-system operation. Together, these findings motivate an interaction-aware view of MAS security: trace attacks end to end, test whether defenses close those paths, and evaluate system-level effects with appropriate counterfactuals.
May 1, 2026

When Embedding-Based Defenses Fail: Rethinking Safety in LLM-Based Multi-Agent Systems

Large language model (LLM)-powered multi-agent systems (MAS) enable agents to communicate and share information, achieving strong performance on complex tasks. However, this communication also creates an attack surface where malicious agents can propagate misinformation and manipulate group decisions, undermining MAS safety. Existing embedding-based defenses aim to detect and prune suspicious agents, but their effectiveness depends on a clear separation between the text embeddings of malicious and benign messages. Attackers can circumvent such defenses by crafting messages whose embeddings lie close to benign ones. We analyze this failure mode theoretically and validate it empirically with three attacks, Slow Drift, Benign Wrapper, and Chaos Seeding. Our analysis further reveals a fundamental limitation of embedding-based defenses: because they rely solely on the text embeddings, they ignore token-level confidence signals such as logits, which can remain informative when embeddings are not distinguishable under attack. We propose using confidence scores to prune or down-weight messages during MAS communication. Experiments show improved robustness across models, datasets, and communication topologies. Moreover, we find that the effectiveness of confidence signals decays over communication rounds, highlighting the importance of early intervention. This insights can inform and inspire future work on MAS attacks and defenses.