attack 2026

RECUR: Resource Exhaustion Attack via Recursive-Entropy Guided Counterfactual Utilization and Reflection

Ziwei Wang 1, Yuanhe Zhang 2, Jing Chen 1, Zhenhong Zhou 3, Ruichao Liang 3, Ruiying Du 1, Ju Jia 4, Cong Wu 5, Yang Liu 3

1 Wuhan University

2 Beijing University of Posts and Telecommunications

3 Nanyang Technological University

4 Southeast University

5 University of Hong Kong

0 citations · 33 references · arXiv (Cornell University)
α

Published on arXiv

2602.08214

Model Denial of Service

OWASP LLM Top 10 — LLM04

Key Finding

RECUR increases LRM output length by up to 11x and decreases throughput by 90% by exploiting the reflective component of chain-of-thought reasoning via Recursive Entropy-guided counterfactual prompts.

RECUR

Novel technique introduced

Large Reasoning Models (LRMs) employ reasoning to address complex tasks. Such explicit reasoning requires extended context lengths, resulting in substantially higher resource consumption. Prior work has shown that adversarially crafted inputs can trigger redundant reasoning processes, exposing LRMs to resource-exhaustion vulnerabilities. However, the reasoning process itself, especially its reflective component, has received limited attention, even though it can lead to over-reflection and consume excessive computing power. In this paper, we introduce Recursive Entropy to quantify the risk of resource consumption in reflection, thereby revealing the safety issues inherent in inference itself. Based on Recursive Entropy, we introduce RECUR, a resource exhaustion attack via Recursive Entropy guided Counterfactual Utilization and Reflection. It constructs counterfactual questions to verify the inherent flaws and risks of LRMs. Extensive experiments demonstrate that, under benign inference, recursive entropy exhibits a pronounced decreasing trend. RECUR disrupts this trend, increasing the output length by up to 11x and decreasing throughput by 90%. Our work provides a new perspective on robust reasoning.

Key Contributions

  • Introduces Recursive Entropy — a metric quantifying the risk of runaway reflection in LRMs, defined as the ratio of a generated token's probability to next-token distribution entropy
  • Proposes RECUR, which constructs counterfactual questions to induce overthinking, then uses Recursive Entropy-guided sampling and coherence-based trimming to distill concise attack prompts that reliably trigger infinite thinking loops
  • Demonstrates empirically that RECUR increases output length by up to 11x and decreases inference throughput by 90% across multiple LRMs

🛡️ Threat Analysis

Details

Domains
nlp
Model Types
llm
Threat Tags
grey_boxinference_timetargeted
Applications
llm inference serviceslarge reasoning model apis
Read PDF arXiv DOI

Similar Papers

attack

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

2026 3 cit.
82%
attack

Rethinking Latency Denial-of-Service: Attacking the LLM Serving Framework, Not the Model

2026 0 cit.
73%
attack

ThinkTrap: Denial-of-Service Attacks against Black-box LLM Services via Infinite Thinking

2025 7 cit.
67%
attack

POT: Inducing Overthinking in LLMs via Black-Box Iterative Optimization

2025 0 cit.
67%
attack

CODE: A Contradiction-Based Deliberation Extension Framework for Overthinking Attacks on Retrieval-Augmented Generation

2026 0 cit.
64%
attack

Asking Forever: Universal Activations Behind Turn Amplification in Conversational LLMs

2026 0 cit.
Model Poisoning
62%
attack

LoopLLM: Transferable Energy-Latency Attacks in LLMs via Repetitive Generation

2025 4 cit.
62%
attack

RepetitionCurse: Measuring and Understanding Router Imbalance in Mixture-of-Experts LLMs under DoS Stress

2025 0 cit.
62%