Unverified50% confidenceFactExact time
Analysis with frozen model weights revealed that hidden text leaves rough signals inside the model, but the glyph identity representation is very weak.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedResearchers 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.71% similarUnverifiedAn engineeringly reliable embedding model should exhibit minimal vector change under character-level and lexical-level perturbations, highly similar vectors under semantic-level paraphrasing, and a perceptible distance change when there is a substantive semantic shift (such as negation).64% similarUnverifiedPrompt injection exploits not a vulnerability in the code parser, but the model's own insufficient awareness of the boundary between instructions and data64% similarUnverifiedModels most easily capture the surface style of Torvalds's emails (straightforward and sharp wording), while the truly scarce deep technical judgment (architectural intuition) is tacit knowledge that is difficult to infer from text.64% similarUnverified模型的格式方言源于指令微调阶段训练数据的格式差异，被模型深度内化形成强烈格式偏好，难以通过运行时提示词完全覆盖63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1001165API
curl https://kongchang.com/api/v1/knowledge/claims/1001165MCP
get_claim(id=1001165)