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 · AWS智能体越权漏洞重大1 分で読めるDisconfirmAI 編集部

Security Firm Zenity Reports Prompt Attacks Can Hijack AWS Agents in the Same Account and Region

Zenity Labs discloses AgentCorruption: a single prompt can steal credentials and control all Bedrock AgentCore agents in an AWS region.

2026-10-08 20:57 GMT+8

Security firm Zenity Labs disclosed "AgentCorruption," a vulnerability chain showing that a single prompt sent to one public-facing Amazon Bedrock AgentCore agent was enough to take control of every agent in the same AWS account and region.

Attackers exploited a lack of network isolation to trick the agent into accessing the Instance Metadata Service (IMDS), stealing temporary AWS credentials. Due to overly broad default permissions, these credentials allowed lateral movement, enabling access to other agents' source code, private conversations, and stored keys, as well as memory manipulation for persistent backdoors.

Zenity reported the issue to AWS in December 2025. AWS subsequently made IMDSv2 the default for new deployments and tightened default execution roles in August, blocking inter-agent invocation and secret retrieval. While partially fixed, researchers recommend enterprises create custom least-privilege roles, noting the inherent conflict between cloud security segmentation and the flexibility AI agents require.

ソース

  • labs.zenity.io↗
トピック

AWS智能体越权漏洞

安全公司Zenity Labs披露了Amazon Bedrock AgentCore平台存在名为“AgentCorruption”的严重安全漏洞链。攻击者仅需向一个公开暴露的AI智能体发送单条提示词,即可通过服务器端请求伪造(SSRF)访问…

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

4 ステップ2 件更新日 2026-10-09 22:19

経緯

  1. 2025-12-25Zenity Labs向AWS负责任地披露了AgentCore中的IMDS访问缺陷及权限配置问题。zenity.io ↗
  2. 2026-04-07Palo Alto Networks Unit 42发布研究,指出AgentCore沙箱网络隔离存在DNS隧道绕过风险,并确认IMDS缺乏会话令牌强制机制。unit42.paloaltonetworks.com ↗
  3. 2026-10-08Zenity Labs正式公开“AgentCorruption”漏洞链详情,展示如何通过单条提示词劫持同区域所有智能体。labs.zenity.io ↗

未解決の問題

  • AWS是否会在未来版本中进一步限制默认角色的跨智能体调用权限,还是继续依赖用户自定义策略?
  • 其他云厂商(如Azure AI Foundry, Google Vertex AI)是否存在类似的IMDS或元数据服务暴露风险?
  • 针对智能体记忆投毒(Memory Poisoning)的自动化检测工具能否有效识别隐蔽的持久化后门?

このスレッド内の他の記事

  • Single Prompt Could Hijack All AI Agents in AWS Region Due to Permission Flaw2026-10-09
トピック全文を読む →
リサーチを続ける →