Unverified50% confidenceOpinionExact time
When the agent text does not match what the user actually says (for example, the user contradicts or interrupts), the STT model needs the ability to suppress this text prior, otherwise it may rely on it excessively.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
Unverified用户可以通过提示注入和语义改写等手段绕过内容检测,使实际拦截效果是概率性的而非绝对的67% similarUnverifiedIf attackers understand the detection mechanism, they may evade detection by manipulating the translation stability of the text.66% similarUnverifiedRefusal and handoff replies use fixed strings rather than model generation, to reduce the attack surface for prompt injection.65% similarUnverifiedReal-time desensitization technology replaces sensitive fields with masks by deploying regular expression matching, named entity recognition models, or dedicated detection models at the STT output layer.65% similarUnverifiedLLM post-processing agents should have strictly constrained prompts that only perform spelling and grammar corrections within the scope of the given transcription text, preventing the completion of information that does not exist.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1133305API
curl https://kongchang.com/api/v1/knowledge/claims/1133305MCP
get_claim(id=1133305)