Terrarium: Revisiting the Blackboard for Multi-Agent Safety, Privacy, and Security Studies

Mason Nakamura, Abhinav Kumar, Saaduddin Mahmud, Sahar Abdelnabi, Shlomo Zilberstein, Eugene BagdasarianarXiv 2025 · DOI · arXiv

Abstract

Terrarium is a new modular framework designed to stress-test the safety, privacy, and security of Large Language Model (LLM) multi-agent systems (MAS). By reviving the classic "blackboard" architecture as a centralized communication proxy, it allows researchers to systematically observe and attack agent collaborations. Our evaluation reveals that even highly capable agents are alarmingly vulnerable to data leaks, coordination sabotage, and low-cost availability attacks, highlighting a critical gap in current MAS safety research.

Introduction

As LLM-based agents move beyond single-user chat to autonomous multi-agent systems—handling tasks like complex meeting scheduling or smart-grid management—they introduce a massive new attack surface. Unlike single-agent setups, MAS rely on inter-agent communication and shared data. Current benchmarks often focus on individual agent performance rather than the systemic risks of collaboration, such as a single malicious agent "poisoning" a group's collective goal or tricking a peer into revealing a user’s private constraints. Terrarium addresses this gap by providing a standardized, "Wireshark-like" environment to study how these interactions fail.

The Terrarium Framework

The authors introduce Terrarium, a framework built on several key architectural choices designed for transparency and control:

  • The Blackboard Design: Instead of direct peer-to-peer messaging, agents interact via "blackboards"—shared workspaces where they post and read messages. This provides a centralized point for researchers to inspect logs, enforce protocols, and inject adversarial interventions.
  • Instruction-Augmented DCOPs: Tasks are framed as Distributed Constraint Optimization Problems (DCOPs). This allows researchers to quantify exactly how much an attack degrades the system's utility using a mathematical "ground truth" global objective.
  • Modular Abstractions: The system decouples agents, environments, tools (using the Model Context Protocol), and communication protocols. This modularity allows researchers to swap out LLMs or attack vectors without rebuilding the entire simulation.
The Terrarium architecture decouples agents, environments, and tools through a centralized communication protocol and blackboard system.

Experimental Setup

The framework is evaluated using three distinct collaborative scenarios based on the Collab benchmark:

  1. Meeting Scheduling: A partially-real-world environment where 10 agents must coordinate 15 meetings across physical locations, accounting for travel time and private user preferences.
  2. Personal Assistant: A synthetic "outfit coordination" task where agents must choose clothes that satisfy both personal style and social constraints.
  3. Smart-Home Assistant: A grid-coordination task using real-world meter data where 8 agents must schedule power-hungry appliances to minimize total peak demand.
An illustration of the Meeting Scheduling task within Terrarium, showing agent preferences, blackboard interactions, and the underlying factor graph for optimization.

Results and Vulnerability Analysis

The authors tested three categories of attacks on the Meeting Scheduling task using GPT-4.1-mini, revealing significant weaknesses in current LLM cooperation:

  • Confidentiality (100% Success): Adversarial agents successfully elicited private information from peers in every single test run, despite system prompts forbidding such disclosure. This suggests that prompting is a weak shield against social engineering between agents.
  • Integrity (Utility Degradation): Communication poisoning (injecting false preferences) successfully lowered the system's joint utility. A clear correlation was observed: as the number of "poisoning shots" increased, the attack efficacy rose.
  • Availability (100% Success): Context overflow attacks—designed to fill an agent's memory window with junk—consistently caused API errors, effectively shutting down coordination at a tiny fraction of the cost required for honest completion.

Conclusion

Terrarium demonstrates that the cooperative bias in modern LLMs makes them easy targets for subtle misinformation and data exfiltration. The "blackboard" acts as a controlled "man-in-the-middle" that makes multi-agent interactions transparent and manipulatable for study. Future work will explore purely competitive environments and the development of automated defense mechanisms, such as "firewall" agents that monitor the blackboard for malicious patterns.

Summary updated

Abstract

A multi-agent system (MAS) powered by large language models (LLMs) can automate tedious user tasks such as meeting scheduling that requires inter-agent collaboration. LLMs enable nuanced protocols that account for unstructured private data, user constraints, and preferences. However, this design introduces new risks, including misalignment and attacks by malicious parties that compromise agents or steal user data. In this paper, we propose the Terrarium framework for fine-grained study on safety, privacy, and security in LLM-based MAS. We repurpose the blackboard design, an early approach in multi-agent systems, to create a modular, configurable testbed for multi-agent collaboration. We identify key attack vectors such as misalignment, malicious agents, compromised communication, and data poisoning. We implement three collaborative MAS scenarios with four representative attacks to demonstrate the framework's flexibility. By providing tools to rapidly prototype, evaluate, and iterate on defenses and designs, Terrarium aims to accelerate progress toward trustworthy multi-agent systems.
BibTeX
@misc{nakamura2025terrarium,
  author = {Mason Nakamura and Abhinav Kumar and Saaduddin Mahmud and Sahar Abdelnabi and Shlomo Zilberstein and Eugene Bagdasarian},
  title = {{Terrarium: Revisiting the Blackboard for Multi-Agent Safety, Privacy, and Security Studies}},
  year = {2025},
  eprint = {2510.14312},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.14312},
}
Figure