Unverified50% confidenceCautionExact time
由于大语言模型存在提示词注入攻击风险和幻觉问题，权限过大的Agent可能在被恶意输入诱导或判断失误时对业务数据、外部系统造成不可逆破坏
1
Sources
50%
Confidence
Long-term
Relevance
7/17/2026
First Seen
Sources
Related Claims
UnverifiedThe autonomous decision-making of agents may produce unpredictable side effects such as accidental code deletion and erroneous deployments, requiring well-developed permission boundaries and behavioral audit mechanisms.78% similarUnverifiedWhen a proactively executing agent makes an error (such as automatically publishing incorrect content), the consequences are more severe than those of a passive tool, placing extremely high demands on alerting and manual confirmation mechanisms.78% similarUnverifiedThe hallucination problem in Agents is more severe than in ordinary dialogue scenarios, because erroneous reasoning is amplified through the action chain.77% similarUnverified在AI安全领域，失准指模型的实际行为偏离设计者或使用者意图的现象，当代大模型的失准更多表现为目标过度追求75% similarUnverifiedThe higher the degree of autonomy of an Agent, the greater the potential data security risk, because the Agent's behavioral paths cannot be fully anticipated at design time.75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/541150API
curl https://kongchang.com/api/v1/knowledge/claims/541150MCP
get_claim(id=541150)