読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
← 読解に戻る
ニューストピック · Mobius注入攻击重大1 分で読めるDisconfirmAI 編集部

New Research: Single Prompt Injection Can Cause Recursive Execution in AI Agents, Amplifying Resource Consumption by 51x

Researchers propose 'Mobius Injection,' an attack paradigm using Semantic Closure to trap AI agents in recursive execution loops, leading to resource-exhaustion DDoS.

2026-10-20 12:00 GMT+8

An arXiv preprint titled "Mobius Injection" introduces a novel attack paradigm that weaponizes autonomous Large Language Model (LLM) agents into zombie nodes to launch Agent-based and -Oriented DDoS (AbO-DDoS) attacks. The study, updated on October 19, 2026, was conducted by researchers including Zi Liang.

While prior research focused on the security of LLMs or agents in isolation, this work exposes the systemic risk of agents acting as disruptive "hubs" between users and services. The core mechanism, termed "Semantic Closure," combines known indirect prompt injection techniques with persistent agent components. A single malicious input triggers sustained recursive execution within the agent. This attack is lightweight for adversaries and evades traditional volumetric DDoS monitors and input filters calibrated for low false-positive rates.

Experiments across three representative claw-style agents and three mainstream coding agents, integrated with 12 frontier proprietary or open-weight LLMs, demonstrated substantial impact. The attack achieved a single-node call amplification factor of up to 51.0x. Under coordinated multi-node pressure, each infected node drained approximately 123k–127k tokens on a shared five-GPU backend. Benign request latency increased by about 23x–24x, eventually saturating shared capacity and effectively denying service to legitimate traffic.

ソース

  • arxiv.org
ログイン / 新規登録
読書リサーチレーダー投資フレームワーク
↗
トピック

Mobius注入攻击

研究人员提出“Mobius Injection”(莫比乌斯注入)攻击范式,揭示了大型语言模型(LLM)智能体作为用户与服务枢纽时的系统性安全风险。该攻击利用“语义闭合”机制,通过单次提示注入即可诱导智能体组件陷入持续递归执行,将正常节点转化…

2 ステップ0 件更新日 2026-10-21 13:54

経緯

  1. 2026-05-12论文《Mobius Injection: From One-Shot Prompt Injection to Sustained Resource Exhaustion in AI Agent Infrastructure》首次提交至arXiv,提出利用语义闭合机制使AI智能体陷入递归执行的攻击方法。arxiv.org ↗
  2. 2026-10-19论文更新至v2版本,详细披露了在三类爪式智能体和三类主流编码智能体上的实验结果,证实单节点调用放大可达51.0倍,并介绍了ACE防御机制。arxiv.org ↗

未解決の問題

  • 目前提出的ACE防御机制在实际生产环境中的误报率和拦截效率如何?
  • 主流LLM服务商是否已针对此类递归注入更新其输入过滤或速率限制策略?
  • 该攻击范式是否会随着智能体框架(如Claude Code, OpenClaw等)的版本迭代而失效或变异?
トピック全文を読む →
リサーチを続ける →