Search:
Match:
4 results

RepetitionCurse: DoS Attacks on MoE LLMs

Published:Dec 30, 2025 05:24
1 min read
ArXiv

Analysis

This paper highlights a critical vulnerability in Mixture-of-Experts (MoE) large language models (LLMs). It demonstrates how adversarial inputs can exploit the routing mechanism, leading to severe load imbalance and denial-of-service (DoS) conditions. The research is significant because it reveals a practical attack vector that can significantly degrade the performance and availability of deployed MoE models, impacting service-level agreements. The proposed RepetitionCurse method offers a simple, black-box approach to trigger this vulnerability, making it a concerning threat.
Reference

Out-of-distribution prompts can manipulate the routing strategy such that all tokens are consistently routed to the same set of top-$k$ experts, which creates computational bottlenecks.

Prompt-Based DoS Attacks on LLMs: A Black-Box Benchmark

Published:Dec 29, 2025 13:42
1 min read
ArXiv

Analysis

This paper introduces a novel benchmark for evaluating prompt-based denial-of-service (DoS) attacks against large language models (LLMs). It addresses a critical vulnerability of LLMs – over-generation – which can lead to increased latency, cost, and ultimately, a DoS condition. The research is significant because it provides a black-box, query-only evaluation framework, making it more realistic and applicable to real-world attack scenarios. The comparison of two distinct attack strategies (Evolutionary Over-Generation Prompt Search and Reinforcement Learning) offers valuable insights into the effectiveness of different attack approaches. The introduction of metrics like Over-Generation Factor (OGF) provides a standardized way to quantify the impact of these attacks.
Reference

The RL-GOAL attacker achieves higher mean OGF (up to 2.81 +/- 1.38) across victims, demonstrating its effectiveness.

Analysis

This article discusses a new type of denial-of-service (DoS) attack, called ThinkTrap, targeting black-box Large Language Model (LLM) services. The attack exploits the LLM's reasoning capabilities to induce an infinite loop of processing, effectively making the service unavailable. The research likely explores the vulnerability and potential mitigation strategies.
Reference

The article is based on a paper published on ArXiv, suggesting a peer-reviewed or pre-print research.

Technology#AI👥 CommunityAnalyzed: Jan 3, 2026 16:09

AI crawlers are overwhelming websites; Meta and OpenAI are the primary culprits

Published:Aug 21, 2025 11:35
1 min read
Hacker News

Analysis

The article highlights a growing problem: the excessive activity of AI crawlers, specifically those from Meta and OpenAI, is causing performance issues and potential denial-of-service for websites. This is a significant concern as it impacts website availability and user experience. The article likely discusses the technical aspects of the problem, such as the volume of requests, the impact on server resources, and potential solutions like rate limiting or bot detection.
Reference