Unverified50% confidenceDecision RuleExact time
If the agent completes the task normally in the benign version but refuses the dangerous call in the malicious version, it is stronger evidence that the security policy is working.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen an agent fails to execute a dangerous operation in the face of a Prompt Injection attack, this may stem from two paths: the security policy catching the malicious path, or the model avoiding the tool or task for unrelated reasons.78% similarUnverifiedWhen authorizing an Agent to perform high-risk operations, a manual confirmation step should be established, where the Agent proposes the plan and humans retain final execution authority (human-in-the-loop mechanism).78% similarUnverifiedThe ideal Agent security solution should adopt declarative security policies that predefine which operations are prohibited, require confirmation, or can be executed freely.78% similarUnverified对于涉及不可逆外部动作的Agent操作应保留人类否决权，让Agent承担推理执行工作而在关键决策节点保留人类确认73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/901660API
curl https://kongchang.com/api/v1/knowledge/claims/901660MCP
get_claim(id=901660)