针对 Claude Computer Use 的间接注入数据破坏 (AML.CS0046)
一句话通俗理解
一份 PDF 就能让 Claude 操作你的电脑「清理」文件——护栏在精心注入面前失守。
案例速览
| 维度 | 内容 |
|---|---|
| 案例类型 | 演练 / 安全研究 |
| 发生时间 | 2024-10-24 |
| 攻击者 | HiddenLayer |
| 受害目标 | Claude Computer Use Agent |
事件经过
HiddenLayer 安全研究者演示了针对 Claude Computer Use 的间接提示词注入:经 PDF 内容注入诱导 Claude 绕过护栏执行 shell 命令删除数据——证明计算机使用类智能体的系统级能力可被文档内容劫持。
运用的 ATLAS 技术
以下攻击链映射来自 MITRE ATLAS 官方数据(AML.CS0046 的
employs关系),按官方步骤编号排序;技术名称链接至本库对应技术文档,引用文字为官方对本次运用的描述。
S00
- LLM 提示词构造 (AML.T0065)
The researchers crafted a prompt targeting Claude’s Computer Use feature. It was designed to bypass and execute a destructive command on the victim’s system.
S01
- 通过公开应用进行提示词渗透 (AML.T0093)
The researchers embedded the malicious prompt in a PDF document. This document could have ended up on the victim’s system through a public-facing application such as email or shared document stores.
S02
- 间接提示词注入 (AML.T0051.001)
When a user asked Claude to interact with the PDF file, the embedded prompt was executed.
S03
- LLM 越狱 (AML.T0054)
The prompt instructed Claude that this is a virtual environment designed for security testing and that it is okay to execute potentially dangerous commands. This bypassed Claude’s guardrails preventing the execution of obfuscated commands.
S04
- LLM 提示词混淆 (AML.T0068)
The malicious command was obfuscated with base64 and rot13 encoding. The prompt included instructions for Claude to decode the command.
S05
- AI 智能体工具调用 (AML.T0053)
Claude Computer Use invoked its
bashtool to execute malicious command.
S06
- 通过 AI 智能体工具调用破坏数据 (AML.T0101)
The shell command executed by Claude Computer Use deleted the victim’s filesystem.
关联战术
- AI 攻击适配 (AML.TA0001)
- 初始访问 (AML.TA0004)
- 执行 (AML.TA0005)
- 权限提升 (AML.TA0012)
- 防御规避 (AML.TA0007)
- 影响 (AML.TA0011)
参考资料
以下为 MITRE ATLAS 官方案例收录的参考来源:
版本: ATLAS 2026.09 | 更新: 2026-10-24 | 模板: ATLAS 案例研究 v1.0