Unverified50% confidenceFactExact time
Term-list words are isolated profanities or abusive terms divorced from sentences, and the model receives only the character sequence with no semantic cues whatsoever, simulating the worst case for short-field moderation; abusive content in comments is mostly embedded within complete sentences, where the model can use collocation, tone, and syntax to judge.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedA model can avoid words or sentence patterns flagged as harmful at the surface output level while retaining the original bias structure in the semantic associations encoded in its weights.78% similarUnverifiedSystem prompt-based solutions are essentially natural language conventions and can easily be bypassed by jailbreak attacks, prompt injection, or unintended model behavior.70% similarUnverifiedUnigram 模型由 SentencePiece 库实现，从大词表出发逐步剔除对整体似然贡献最小的子词来缩减词表69% similarUnverifiedResearchers have found that hidden character techniques can be used to launch prompt injection attacks against large language models, embedding hidden instructions within normal text that the model will parse and execute.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1015397API
curl https://kongchang.com/api/v1/knowledge/claims/1015397MCP
get_claim(id=1015397)