Unverified50% confidenceFactExact time
When reading external documents, an AI assistant cannot reliably distinguish between text being summarized and instructions, and if it has tool permission to send email, hidden instructions may be executed.
1
Sources
50%
Confidence
Long-term
Relevance
10/18/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedFor AI agents, any content they read can potentially be interpreted as instructions, blurring the boundary between data and instructions.78% similarUnverifiedAI Agent无法将PRD文档写入磁盘的根本原因是工具权限配置中缺少写入权限，而非提示词不够清晰75% similarUnverifiedWhen processing ambiguous instructions, AI cannot proactively identify that a given response is unsuitable as a file name, nor will it ask the user what they want the file to be called.73% similarUnverifiedThe conversation record formats of different AI coding assistant tools are mutually incompatible, each sealed within its own isolated data silo73% similarUnverifiedUsing AI to draft emails is itself normal, but the content sent must be accurate and something one is willing to take responsibility for; if there are company rules, one should comply with restrictions on the use of AI tools or the pasting of sensitive information.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1063948API
curl https://kongchang.com/api/v1/knowledge/claims/1063948MCP
get_claim(id=1063948)