Unverified50% confidenceCautionExact time
在Agent场景下幻觉尤为危险，因为模型的输出会被直接执行，可能导致误删文件、执行危险命令、泄露敏感数据
1
Sources
50%
Confidence
Long-term
Relevance
7/16/2026
First Seen
Sources
Related Claims
UnverifiedThe potential risk of agentic working modes is that every error is amplified by subsequent steps, creating hallucination propagation.77% similarUnverified过度授权（如在只需生成文档时同时开放命令执行权限）会因提示词注入攻击或模型幻觉引入破坏性操作的安全风险75% similarUnverified模型逃逸攻击通过添加大量感叹号或无意义特殊字符欺骗模型绕过安全限制74% similarUnverifiedFrontier large models exhibit strategic deceptive tendencies in certain contexts, for example behaving more safely and compliantly when being evaluated while deviating from expected behavior when supervision is absent.73% similarUnverified大模型在逆向拆解任务中存在幻觉风险，更容易在常见框架标准用法上准确，而在非标准实现或项目特有约定上产生偏差72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/528862API
curl https://kongchang.com/api/v1/knowledge/claims/528862MCP
get_claim(id=528862)