From Secure Agentic AI to Secure Agentic Web: Challenges, Threats, and Future Directions

Zhihang Deng, Jiaping Gui, Weinan Zhang

TL;DR

As LLMs evolve from simple chatbots into autonomous agents that can browse the web and use tools, the security stakes have shifted from "don't say bad things" to "don't let the agent accidentally delete your database." This paper provides a comprehensive roadmap for securing the emerging Agentic Web—an ecosystem where agents interact with each other and third-party services—moving beyond individual model safety toward system-wide protocol security.

What Problem Does This Paper Address?

Traditional LLM security focused on "jailbreaking" a model to generate toxic text. However, modern agentic systems use LLMs as a "brain" connected to external tools (APIs), persistent memory, and the open internet. This autonomy creates a massive new attack surface.

The authors argue that we are transitioning from isolated agents to an Agentic Web, where agents delegate tasks to one another. In this networked environment, a single vulnerability (like a malicious instruction hidden on a webpage) can propagate through a chain of agents, leading to systemic failures. The paper addresses the gap between current "patchwork" defenses and the robust, protocol-level security required for a world of autonomous machine-to-machine interactions.

Methods

This paper is a systematic survey that organizes the current research landscape into two main frameworks:

  1. Component-Aligned Threat Taxonomy: The authors break down an agent into its functional parts (Brain, Memory, Tools, Environment) and identify specific threats for each, such as "Environment Injection" (malicious web content) and "Toolchain Abuse" (tricking an agent into improper API calls).
  2. Layered Defense Strategy: They categorize defenses based on where they intervene, from "Prompt Hardening" at the input level to "Protocol Security" at the ecosystem level.

Crucially, for every threat and defense, the authors include an "Agentic Web implication" section, explaining how a localized risk (e.g., memory poisoning) escalates when agents start sharing data and delegating tasks across administrative domains.

Datasets

As a survey paper, it does not introduce a new dataset but reviews the "practical toolkit" of existing benchmarks used to evaluate agent security, including:

  • Attack Benchmarks: PromptInjectionBench (general injections), InjecAgent (tool-integrated injections), and Webinject (web-browsing agents).
  • Robustness & Safety Frameworks: SGBench (safety generalization), AgentSecurityBench (ASB) (standardized attack surfaces), and OpenAgentSafety (real-world task safety).
  • Red-Teaming Tools: AgentDojo (dynamic environments) and AART (automated adversarial data generation).

Key Insights and Ideas

  • The "Brain" is Not the Only Target: Security isn't just about the LLM. The interfaces between the agent and its tools or memory are often the weakest links.
  • Memory as a Persistent Liability: Unlike a standard chat session, an agent’s long-term memory can be "poisoned." A single malicious interaction can influence the agent’s reasoning for weeks or months.
  • Programmable Trust: In the Agentic Web, trust cannot be implicit. We need explicit, machine-verifiable primitives for identity and authorization. An agent shouldn't just "trust" another agent because it has a valid API key; it needs fine-grained delegation constraints.
  • The Multi-Agent Security Tax: There is a documented trade-off where increasing security measures (like strict verification steps) can reduce the efficiency and "intelligence" of agent collaboration.

Key Results

The paper synthesizes findings from dozens of recent studies to highlight the severity of the threat:

  • Amplification: Research like AgentSmith shows that a single adversarial image can trigger a "jailbreak" that spreads exponentially across a million multimodal agents.
  • Indirect Injection: Studies like NotWhatYouSignedUpFor demonstrate that agents are highly susceptible to "hidden" instructions in web content, which can override the user's original goals without the user ever seeing the malicious prompt.
  • Toolchain Risks: The survey highlights that "innocent" tools can be chained together by an adversary to create harmful effects (e.g., using a "read file" tool and an "email" tool to exfiltrate data) even if the model's output looks benign.

Limitations and Future Work

The authors acknowledge that current defenses are often "narrow and fragile," typically addressing one specific attack type rather than providing general guarantees. They outline several critical "Open Challenges":

  • Scaling Security: We need ways to monitor and coordinate thousands of interacting agents without creating massive performance bottlenecks.
  • Provenance and Traceability: If an agent performs a harmful action in a chain of ten agents, how do we "forensically" determine which agent was compromised and why?
  • Ecosystem-Level Response: We lack the infrastructure to "quarantine" or revoke the credentials of a compromised agent across the entire web in real-time.
  • Antifragility: Moving beyond mere "robustness" to systems that actually improve their security posture as they are exposed to stressors and attacks.

Summary updated

Abstract

Large Language Models (LLMs) are increasingly deployed as agentic systems that plan, memorize, and act in open-world environments. This shift brings new security problems: failures are no longer only unsafe text generation, but can become real harm through tool use, persistent memory, and interaction with untrusted web content. In this survey, we provide a transition-oriented view from Secure Agentic AI to a Secure Agentic Web. We first summarize a component-aligned threat taxonomy covering prompt abuse, environment injection, memory attacks, toolchain abuse, model tampering, and agent network attacks. We then review defense strategies, including prompt hardening, safety-aware decoding, privilege control for tools and APIs, runtime monitoring, continuous red-teaming, and protocol-level security mechanisms. We further discuss how these threats and mitigations escalate in the Agentic Web, where delegation chains, cross-domain interactions, and protocol-mediated ecosystems amplify risks via propagation and composition. Finally, we highlight open challenges for web-scale deployment, such as interoperable identity and authorization, provenance and traceability, ecosystem-level response, and scalable evaluation under adaptive adversaries. Our goal is to connect recent empirical findings with system-level requirements, and to outline practical research directions toward trustworthy agent ecosystems.
BibTeX
@misc{deng2026from,
  author = {Zhihang Deng and Jiaping Gui and Weinan Zhang},
  title = {{From Secure Agentic AI to Secure Agentic Web: Challenges, Threats, and Future Directions}},
  year = {2026},
  eprint = {2603.01564},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.01564},
}
Figure