← Back to Paper Radar

Beyond Max Tokens: Stealthy Resource Amplification via Tool Calling Chains in LLM Agents

Kaiyu Zhou, Yongsen Zheng†, Yicheng He, Meng Xue, Xueluan Gong, Yuji Wang, Kwok-Yan Lam
arXiv.org | 2026
A malicious but protocol-compatible MCP tool server can mount a stealthy, multi-turn economic denial-of-service attack on LLM agents. It edits only text-visible fields and a template-governed return policy, optimized via MCTS, so that the agent is driven into long, verbose tool-calling chains while the final answer stays correct.

Problem Statement

Existing LLM DoS attacks, such as prompt-based or RAG-injection sponge attacks, are single-turn and bounded by per-response max-token caps. They are often not task-oriented, which makes them conspicuous in goal-driven agent workflows. They also fail to exploit the compounding cost of multi-turn agent-tool loops, so the agent-tool interface is an under-examined attack surface where output-correctness checks give a false sense of safety.

Key Novelty

  • A tool-layer, multi-turn economic DoS threat model. The attacker controls a benign-looking MCP-compatible tool server, leaves function signatures unchanged, and preserves the final payload, so the task still completes correctly.
  • Cost amplification that escapes single-turn max-token caps by compounding tokens across many agent-tool turns, triggered only through text-only notices in tool responses.
  • A Monte Carlo Tree Search optimizer that searches over edits to text-visible tool fields and the return-policy template to maximize trajectory length and cost while staying stealthy.
  • Systems-level impact analysis covering energy, GPU KV-cache occupancy, and co-tenant throughput degradation, beyond token counts alone.

Evaluation Highlights

  • Across six LLMs on ToolBench and BFCL, attacked tasks expand into trajectories exceeding 60,000 tokens, with cost inflation up to 658x and energy increases of 100-560x.
  • GPU KV-cache occupancy rises from <1% to 35-74%, and the throughput of co-running workloads drops by roughly 50%, all while task outcomes remain correct and conventional checks fail.

Signal Assessment

7/10 The paper identifies a practically important attack surface: the MCP/tool layer and multi-turn cost compounding. It demonstrates very large, stealthy resource amplification with real serving-infrastructure impact, which shifts the focus of agent security toward process-level cost monitoring. The underlying ideas, sponge attacks and tool poisoning, extend prior lines of work rather than creating a wholly new paradigm.

Methodology

  1. Define the threat model: an attacker-controlled, MCP-compatible tool server that must keep function signatures and the final correct payload intact so the attack evades output validation.
  2. Parameterize the attack surface as the text-visible fields (e.g., descriptions, notices) and a template-governed return policy that dictates how and when the tool returns partial results or asks for further calls.
  3. Use MCTS to search over edits to these fields and templates, rewarding long, verbose multi-turn tool-calling trajectories that still end in a correct answer.
  4. Deploy the optimized server against six LLM agents on ToolBench and BFCL tasks, measuring tokens, monetary cost, energy, KV-cache occupancy, and co-tenant throughput.

System Components

Malicious MCP-compatible tool server

A benign-appearing tool that conforms to the protocol and keeps signatures unchanged but embeds text-only notices that induce extra calls.

Template-governed return policy

Controls how responses are staged across turns, e.g., withholding or segmenting the final payload, to prolong the interaction while ultimately delivering the correct result.

Text-visible field edits

Modified tool descriptions and response text that steer the agent's reasoning toward verbose, repeated tool invocations.

MCTS optimizer

Explores the combinatorial space of field and template edits to maximize cost amplification under stealth and correctness constraints.

Resource impact measurement

Quantifies token and cost inflation, energy usage, GPU KV-cache occupancy, and throughput degradation for co-located workloads.

Results

Metric/Benchmark Baseline This Paper Delta
Trajectory length (ToolBench/BFCL) Normal short task trajectory >60,000 tokens Large multi-turn expansion
Monetary cost 1x (benign tool) Up to 658x Up to +658x
Energy consumption 1x 100-560x +100-560x
GPU KV-cache occupancy <1% 35-74% +35-73 pts
Co-running throughput 100% (unaffected) ~50% ~-50%
Task correctness / validation Correct Correct (preserved) No detectable change
Prior single-turn DoS (prompt/RAG) Capped by max tokens, conspicuous Escapes caps, stealthy Qualitatively stronger

Key Takeaways

  • Final-answer validation is insufficient for agent security. Monitor per-task token budgets, turn counts, and tool-call frequency, and alert on anomalous cost trajectories.
  • Treat third-party MCP servers as untrusted. Even signature-compatible tools can steer agents through descriptions and response text, so vet tool metadata, sandbox responses, and consider stripping or summarizing tool-provided instructions.
  • Enforce agent-level resource limits (max turns, cumulative token and cost caps, per-tool call quotas), not just per-response max_tokens.
  • Multi-tenant serving operators should account for KV-cache exhaustion from long agent trajectories, which can halve co-tenant throughput. Admission control and per-session KV quotas are relevant mitigations.

Abstract

The agent-tool communication loop is a critical attack surface in modern Large Language Model (LLM) agents. Existing Denial-of-Service (DoS) attacks, primarily triggered via user prompts or injected retrieval-augmented generation (RAG) context, are ineffective for this new paradigm. They are fundamentally single-turn and often lack a task-oriented approach, making them conspicuous in goal-oriented workflows and unable to exploit the compounding costs of multi-turn agent-tool interactions. We introduce a stealthy, multi-turn economic DoS attack that operates at the tool layer under the guise of a correctly completed task. Our method adjusts text-visible fields and a template-governed return policy in a benign, Model Context Protocol (MCP)-compatible tool server, optimizing these edits with a Monte Carlo Tree Search (MCTS) optimizer. These adjustments leave function signatures unchanged and preserve the final payload, steering the agent into prolonged, verbose tool-calling sequences using text-only notices. This compounds costs across turns, escaping single-turn caps while keeping the final answer correct to evade validation. Across six LLMs on the ToolBench and BFCL benchmarks, our attack expands tasks into trajectories exceeding 60,000 tokens, inflates costs by up to 658x, and raises energy by 100-560x. It drives GPU KV cache occupancy from<1% to 35-74% and cuts co-running throughput by approximately 50%. Because the server remains protocol-compatible and task outcomes are correct, conventional checks fail. These results elevate the agent-tool interface to a first-class security frontier, demanding a paradigm shift from validating final answers to monitoring the economic and computational cost of the entire agentic process.

Generated automatically from available metadata and abstract on 2026-11-09.