Sponge Tool Attack: Stealthy Denial-of-Efficiency against Tool-Augmented Agentic Reasoning

Qi Li, Xinchao Wang

As artificial intelligence progresses toward Artificial General Intelligence (AGI), Large Language Models (LLMs) are evolving from passive conversationalists into active agents. By integrating external tools—such as search engines, calculators, and code interpreters—these agentic systems can execute complex reasoning tasks. However, researchers from the National University of Singapore have uncovered a novel, stealthy vulnerability in this paradigm: the "Sponge Tool Attack" (STA). This attack forces agents to "soak up" massive amounts of computational resources and API costs without alerting the user or provider.

The Core Problem: Denial-of-Efficiency (DoE)

Most LLM security research centers on "jailbreaking" (bypassing safety filters) or "poisoning" (corrupting training data). As LLMs become agentic, a new bottleneck arises: efficiency.

Tool invocations are computationally expensive and time-consuming. The researchers introduce a Denial-of-Efficiency (DoE) threat, where an adversary aims to degrade service or bankrupt the provider rather than break the model's safety guardrails. Crucially, STA operates under a strict "query-only" constraint—the attacker has no access to the model's weights or underlying code and can only manipulate the input text.

Methodology: How the Sponge Tool Attack Works

STA operates as an iterative, multi-agent collaborative framework designed to find the most "absorbent" prompt rewrites. It functions in two main phases:

  1. Offline Policy Bank Construction: The system learns how to induce inefficiency through a multi-agent setup. A Prompt Rewriter attempts to complicate a benign query. A Quality Judge evaluates the rewrite, rewarding it for triggering extra tool steps (the "bonus") while penalizing it for altering the original meaning. Finally, a Policy Inductor distills successful attempts into generalized rewriting policies (e.g., "Add a verification layer" or "Force explicit step ordering").
  2. Query-Aware Sponging: When a live query is received, the system selects the most appropriate policy from the bank and rewrites the prompt before sending it to the victim agent.

To maintain stealth, the attack avoids explicitly naming tools and ensures the final answer remains accurate. To the user, the agent simply appears to be exceptionally thorough.

Key Insights

  • The Capability Paradox: Counterintuitively, stronger agentic frameworks are more vulnerable. As agents become more adept at following complex instructions and utilizing tools, they are more easily led down convoluted, inefficient reasoning trajectories.
  • Error Accumulation: The attack amplifies "reasoning noise." By unnecessarily forcing the model into extended reasoning stages, it increases the likelihood of the model misselecting a tool, triggering a runaway loop of tool invocations.
  • Model-Agnostic Transferability: The rewriting policies are highly transferable. A policy learned on a smaller open-source model can effectively "sponge" advanced closed-source APIs like GPT-4.
  • Stealth via Semantic Fidelity: By utilizing embedding-based similarity scores, STA ensures the rewritten prompt retains the original semantic intent, making it highly evasive against standard anomaly detectors.

Results and Impact

The researchers conducted extensive evaluations across 13 datasets spanning five domains (General, Math, Science, Medical, and Agentic), testing 6 models and 4 agentic frameworks.

  • Cost Inflation: The attack drastically increased API latency and monetary costs, inflating expenses by up to 3.4x on datasets like VQAv2.
  • Increased Overhead: STA consistently forced agents to hit their maximum "step budget" (a high "Cap Hit" rate) by inducing unnecessary tool-calling loops.
  • Maintained Accuracy: Despite the bloated reasoning process, task accuracy remained within 1-2% of the original baseline, proving the attack disrupts efficiency without breaking functionality.
  • Human Validation: In human-in-the-loop evaluations, experts agreed that the rewritten prompts preserved the original user intent 73-79% of the time.

Conclusion and Defensive Outlook

The Sponge Tool Attack highlights a critical blind spot in the deployment of LLM agents. While the current iteration relies on an offline phase to build its policy bank, future variants could generate policies on-the-fly. For developers, this research is a clear wake-up call: securing agentic systems requires moving beyond traditional safety guardrails to implement cost-aware monitoring and anomaly detection for tool-call rates, ensuring these systems cannot be quietly squeezed dry.

Summary updated

Abstract

Enabling large language models (LLMs) to solve complex reasoning tasks is a key step toward artificial general intelligence. Recent work augments LLMs with external tools to enable agentic reasoning, achieving high utility and efficiency in a plug-and-play manner. However, the inherent vulnerabilities of such methods to malicious manipulation of the tool-calling process remain largely unexplored. In this work, we identify a tool-specific attack surface and propose Sponge Tool Attack (STA), which disrupts agentic reasoning solely by rewriting the input prompt under a strict query-only access assumption. Without any modification on the underlying model or the external tools, STA converts originally concise and efficient reasoning trajectories into unnecessarily verbose and convoluted ones before arriving at the final answer. This results in substantial computational overhead while remaining stealthy by preserving the original task semantics and user intent. To achieve this, we design STA as an iterative, multi-agent collaborative framework with explicit rewritten policy control, and generates benign-looking prompt rewrites from the original one with high semantic fidelity. Extensive experiments across 6 models (including both open-source models and closed-source APIs), 12 tools, 4 agentic frameworks, and 13 datasets spanning 5 domains validate the effectiveness of STA.
BibTeX
@inproceedings{li2026sponge,
  author = {Qi Li and Xinchao Wang},
  title = {{Sponge Tool Attack: Stealthy Denial-of-Efficiency against Tool-Augmented Agentic Reasoning}},
  booktitle = {International Conference on Machine Learning},
  year = {2026},
  url = {https://openreview.net/forum?id=g9fZTb53kL},
}
Figure