Security of Internet of Agents: Attacks and Countermeasures

Yuntao Wang, Yanghe Pan, Shaolong Guo, Zhou Su

TL;DR

As LLMs evolve from chatbots into autonomous agents, we are witnessing the birth of the Internet of Agents (IoA)—a decentralized ecosystem where AI entities collaborate to solve complex tasks. This survey provides the first comprehensive security roadmap for the IoA, identifying unique "agentic" threats like hallucination cascades and cross-domain safety misalignments, while proposing a multi-layered defense strategy ranging from semantic-aware authentication to multi-agent debate.

What Problem Does This Paper Address?

Traditional cybersecurity focuses on protecting human users and static data. However, the rise of "agentic AI" (autonomous systems powered by LLMs) introduces a paradigm shift. In the IoA, agents negotiate via natural language, share latent embeddings, and autonomously form collaboration networks.

The authors argue that existing security frameworks are ill-equipped for this "agentic web." They address the gap between single-model safety and the systemic risks that emerge when heterogeneous agents interact. Specifically, they tackle how decentralized coordination, semantic ambiguity in protocols, and the coupling of virtual reasoning with physical "embodied" actions (like drones or robots) create a massive, unexplored attack surface.

Methods

This paper is a systematic survey and taxonomy. Rather than introducing a single algorithm, the authors synthesize recent literature to define the IoA architecture and categorize its threats into four critical pillars:

  1. Identity Authentication: Moving beyond passwords to handle "Intent Deception" and "Identity Forgery" (e.g., using AI-generated voice/face to bypass auth).
  2. Cross-Agent Trust: Addressing how malicious or "hallucinating" agents can corrupt a collective workflow.
  3. Embodied Security: Focusing on the "cyber-physical" link, where a prompt injection in the cloud can lead to a physical crash in the real world.
  4. Privacy: Examining how agents might inadvertently "memorize" and leak sensitive user data during multi-turn collaborations.

The paper also reviews emerging communication protocols like Anthropic’s Model Context Protocol (MCP) and Google’s Agent-to-Agent (A2A) protocol, evaluating their built-in security features.

Datasets

As a survey paper, it does not introduce a new experimental dataset. Instead, it references and categorizes results from dozens of primary research papers and benchmarks, such as:

  • RAG-Thief: For evaluating privacy leaks in Retrieval-Augmented Generation.
  • BadRobot: A benchmark for jailbreaking embodied LLMs.
  • PoisonedRAG: For testing knowledge base contamination.

Key Insights and Ideas

  • The Hallucination Cascade: This is a "network effect" for errors. A small hallucination by a retrieval agent can be amplified by a reasoning agent, leading to a "snowball effect" that results in systemic failure across the entire agent network.
  • Intent Deception: Unlike traditional malware, a malicious agent might act perfectly "helpful" during the initial authentication stage to gain access to a government or corporate network, only to pivot to "attack mode" once it has cleared the perimeter.
  • Cross-Domain Safety Misalignment: The authors highlight a fascinating "linguistic vs. action" gap. An embodied agent might correctly refuse a harmful request in text (e.g., "I cannot hurt humans") while simultaneously generating the executable robot code to perform that very action.
  • Semantic-Aware Defenses: Because IoA protocols are often based on natural language, static firewalls are useless. The paper advocates for "multi-agent debate" as a defense mechanism, where agents cross-verify each other's outputs to detect anomalies.

Key Results

The paper provides a definitive Taxonomy of IoA Threats (see Figure 1 in the paper), mapping specific attack vectors to their corresponding defenses:

Threat Category Key Attack Example Proposed Defense
Identity Foice [6]: Cross-modal voice forgery. Decentralized Identifiers (DIDs) & Verifiable Credentials.
Trust Pandora [31]: Multi-step jailbreak via phishing agents. Multi-agent debate & RAG grounding.
Embodied Contextual Backdoors [10]: Triggers hidden in physical objects. World models & multimodal consistency checks.
Privacy Agent Memorization [55]: Leaking PII from training data. Privacy pre-assessment & output filtering.

The authors conclude that decentralization is both a strength (resilience) and a weakness (amplifies Sybil attacks), requiring a "Zero Trust" approach specifically tailored for agentic reasoning.

Limitations and Future Work

The authors acknowledge several open challenges:

  • Low-Latency Revocation: In a decentralized IoA, quickly revoking the "credentials" of a compromised agent across millions of nodes is currently an unsolved engineering problem.
  • Formal Verification: Most current agent protocols are flexible but lack the formal mathematical proofs required for mission-critical safety.
  • Ethical "Justification Logs": Future agents need to provide not just an output, but a "trace" of their reasoning that can be audited by humans to ensure ethical compliance.
  • Cloud-Edge Coordination: Balancing the heavy compute needs of LLM agents with the real-time requirements of edge devices (like autonomous vehicles) remains a significant bottleneck.

Summary updated

Abstract

With the rise of large language and vision-language models, AI agents have evolved into autonomous, interactive systems capable of perception, reasoning, and decision-making. As they proliferate across virtual and physical domains, the Internet of Agents (IoA) has emerged as a key infrastructure for enabling scalable and secure coordination among heterogeneous agents. This survey offers a comprehensive examination of the security and privacy landscape in IoA systems. We begin by outlining the IoA architecture and its distinct vulnerabilities compared to traditional networks, focusing on four critical aspects: identity authentication threats, cross-agent trust issues, embodied security, and privacy risks. We then review existing and emerging defense mechanisms and highlight persistent challenges. Finally, we identify open research directions to advance the development of resilient and privacy-preserving IoA ecosystems.
BibTeX
@article{wang2025security,
  author = {Yuntao Wang and Yanghe Pan and Shaolong Guo and Zhou Su},
  title = {{Security of Internet of Agents: Attacks and Countermeasures}},
  journal = {IEEE Open Journal of the Computer Society},
  year = {2025},
  doi = {10.1109/ojcs.2025.3589638},
}

Loading related content...

Figure