Unverified50% confidenceFactExact time
工具调用机制中，模型不会直接执行任何代码，而是返回结构化的调用请求，由本地程序执行后再回传结果；因此安全边界掌握在开发者手里，模型只能请求，不能擅自读取文件或发送邮件。
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAfter receiving a request, the model may return a structured tool call rather than directly providing an answer.80% similarUnverifiedThe model never directly calls code; it simply returns a message requesting that a tool be run, and it is the loop written by the developer that actually decides what gets executed.76% similarUnverifiedClosed-source models cannot be fine-tuned and users can only influence outputs through prompt engineering, whereas open-weight models allow enterprises to download weights and fine-tune them on their own compute.74% similarUnverifiedExternal text reaching the model (emails, web pages, webhooks) may carry injected instructions targeting the model, so model output should never directly trigger deletions, sends, or payments—a whitelist or code check should be placed in between.73% similarUnverifiedIn the "LLM + tool calling" paradigm, the model does not directly manipulate files itself, but instead outputs structured instructions, with deterministic code handling the actual execution.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1051511API
curl https://kongchang.com/api/v1/knowledge/claims/1051511MCP
get_claim(id=1051511)