Unverified60% confidenceFactTime unknown
Indirect Prompt Injection attacks embed instructions in external data processed by the AI, and since 2023 security researchers have demonstrated this type of attack in multiple AI products.
2
Sources
60%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Related Entities
Related Claims
Unverified研究人员已发现针对AI代码建议的提示注入、通过恶意仓库影响模型输出的数据投毒，以及利用Copilot上下文窗口实施的信息泄露攻击等新型攻击路径79% similarVerifiedThe attack exploits the automatic execution characteristics of AI agents to conceal malicious behavior within normal operational workflows, bypassing static code scanning tools.77% similarUnverified2024 年多家安全研究机构发现，自主 AI Agent 在执行文件操作时可能因上下文理解偏差导致误删关键文件，或在依赖安装过程中引入存在漏洞的第三方包76% similarUnverifiedAdversarial attacks involve applying imperceptibly small perturbations to input data that cause AI models to make erroneous judgments, and have been demonstrated to be feasible in both image recognition and text processing tasks.75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/25198API
curl https://kongchang.com/api/v1/knowledge/claims/25198MCP
get_claim(id=25198)