Unverified50% confidenceDecision RuleExact time
The security boundary of tool calling should rely on code performing checks before execution (parameter validation, permission whitelists, manual confirmation for high-risk actions), not on prompts constraining the model.
1
Sources
50%
Confidence
Long-term
Relevance
11/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSecurity should exist in the execution path (approval binding, state version checks, idempotency keys) rather than in the prompt; a policy that a model can talk its way around does not deserve to be called a guardrail.79% similarUnverifiedGuards sit between the model and the tool functions, bearing responsibility for safety validation by checking before any tool is called, and are a key mechanism for defending against prompt injection and abnormal operations.79% similarUnverifiedThe pre-tool-call check must be executed inline at the tool boundary—meaning the decision is made before the call, not after the result is returned.78% similarUnverified安全检测层往往在代码进入沙箱执行之前就基于代码外表而非实际危害进行分析,导致误报78% similarUnverifiedThe security layer must be placed after tool call parsing and before execution.78% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1177738API
curl https://kongchang.com/api/v1/knowledge/claims/1177738MCP
get_claim(id=1177738)