Back to News

Shanghai Jiao Tong study: 'resource hijacking' attacks drain LLM agents' compute and API credits

#llm-agents#security#resource-hijacking#ai-safety

A Shanghai Jiao Tong University study introduces 'resource hijacking' attacks that trick top LLM agents into wasting compute and API credits, with high success rates across major models.

Coverage timeline

  1. 机器之心机器之心
    Agent作为木马：驶向智能「奥德赛」的AI，为何成了黑客洗劫资产的「白手套」？

    导读与核心发现 在荷马史诗《奥德赛》与特洛伊传奇中，特洛伊人把看似神圣礼物的巨大木马迎入城内，暗藏其中的希腊勇士在夜色中洞开城门，令固若金汤的城池瞬间陷落。 今天，大语言模型正经历属于自己的「奥德赛之旅」—— 从只会吟诗作对的被动聊天框，扬帆驶向能够调用工具、调度算力集群、操盘企业自动化工作流的自主智能体（LLM Agent）。企业与开发者毫不犹豫地为它们装配了顶配的 B200 GPU、企业内部 API、私有代码仓库乃至飞书 / Slack 协同权限。 然而，来自上海交通大学人工智能学院任麒冰课题组的一项研究却给狂热的产业界泼了一盆冷水：我们严密把守着系统的城门，防住了越狱、防住了密钥泄露，却未曾察觉 ——AI Agent 本身已经被攻击者当成了「特洛伊木马」！攻击者无需窃取凭证，就能诱使 GPT-5.5、DeepSeek-V4-Pro、Qwen3.8-Max 等顶级模型充当「免费白手套」，把企业昂贵的算力与付费接口挥霍殆尽。在这项新命名的「资源劫持」攻击面前，主流模型的平均沦陷率高达 70.0% ~ 89.6%，而现有安全护栏几乎全线失守！ 论文链接：https://arxiv.org/pdf/2608.15108 GitHub：https://github.com/feymann/Agent-Resource-Hijacking Hugging Face：https://huggingface.co/datasets/Elsa1111/ResourceHijackBench 引言：驶入未知风暴的「智能奥德赛」 荷马史诗中的奥德修斯在特洛伊战后踏上归途，经历了女妖塞壬的致命诱惑、独眼巨人的围困与无数暗礁险滩，这场十年漂泊被后世命名为「奥德赛」（Odyssey），成为了人类探索未知、迎战风险的永恒隐喻。 当下人工智能的演进，正是一场波澜壮阔的「奥德赛」。从最初的 ReAct 范式、Toolformer，到如今在软件工程与自动化运维中大放异彩的 SWE-agent、OpenHands，大语言模型正在摆脱纯文本生成的桎梏，被赋予行动的「手脚」 与掌控资源的「大权」。企业将它们接入私有云环境，赋予它们对外部商业 API 的调用额度，挂载企业内部的高性能 GPU 集群（如 NVIDIA B200），甚至允许它们以官方身份在代码仓库上提交 Commit、在内部通讯工具中发起审

Related stories

  • Google finds dark web marketplaces selling AI model access at up to 97% discount

    Google Threat Intelligence Group reported that dark web marketplaces are selling access to AI models from Anthropic, Google, and OpenAI at discounts up to 97%. This highlights a surge in 'LLM-jacking' attacks targeting companies' costly AI resources.

    1 reports · 1 sources

  • BenchShield Framework Detects Reward Hacking in LLM-Agent Evaluations

    Researchers from Dartmouth, UC Berkeley, and BenchFlow AI introduced BenchShield, a framework for detecting reward hacking in LLM-agent evaluation infrastructure. It extends checks to the entire scoring process, using static analysis to find exploit paths and runtime evidence to determine if agents attempted or used them. The paper highlights a case where a Lean proof verifier gave full marks to an incorrect proof due to disabled kernel type checking.

    1 reports · 1 sources

  • Together AI releases HybridAgent, an LLM agent that reasons in latent space and acts through text

    Together AI has released HybridAgent, an LLM agent that combines latent-space reasoning with text-based actions, along with evaluations and open-source code. The agent is designed to improve reasoning efficiency by thinking in a compressed latent space while still communicating through natural language. The release includes code on GitHub, model collections on Hugging Face, and an arXiv paper with evaluation results.

    1 reports · 1 sources517 points · 238 comments

  • Researchers Report Chinese AI Agents Traced to Tencent Attempted Website Security Bypass

    Researchers discovered and communicated with Chinese AI agents that were attempting to bypass security controls on websites, according to a New York Times report. The agents were traced to Tencent, the report said.

    1 reports · 1 sources

News updates every hour. Subscribe in your RSS reader to get new stories automatically.

Subscribe