Unverified50% confidenceFactExact time
Prompt injection cannot be fully filtered, only contained, because it is hidden instructions buried in the content an Agent reads, and the model treats them like user instructions.
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedConfiguring disable model invocation: true prevents the model from actively invoking the Skill; it can only be triggered manually by the user via a slash command, and the content will not be sent to the model.66% similarUnverified将Codex权限设置为"完全允许"会使Agent在整个文件系统范围内的读写与命令执行绕过人工审核，若封装包含恶意指令集损害将在用户无察觉下完成65% similarUnverifiedA Prompt Handler is essentially a single tool-free LLM call that can only respond based on the text content it receives and cannot actively read files or execute commands.64% similarUnverifiedSystem Prompt与用户输入本质上都是模型的输入token，模型并不存在独立的规则执行引擎来强制保证指令被遵守64% similarUnverifiedThe screen-reading approach isolates the authentication step within manual operation, so the AI can only see pages the user actively opens and cannot initiate requests, turn pages, or export data on its own.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1052820API
curl https://kongchang.com/api/v1/knowledge/claims/1052820MCP
get_claim(id=1052820)