Unverified50% confidenceCautionExact time
Special control tokens (such as <|endoftext|>), if injected through user input, may bypass the safety boundaries of a dialogue system.
1
Sources
50%
Confidence
Long-term
Relevance
9/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLlama 2 的 chat 模板使用 [INST] 和 [/INST] 标记界定用户指令，使用 <<SYS>> 和 <</SYS>> 标记界定系统提示68% similarUnverifiedDesign patterns for handling dangerous operations in conversational interfaces include multi-step confirmation mechanisms, operation previews, time window restrictions, and interactive message components.66% similarUnverifiedAttackers tamper with a model's chat template by calling the create API, permanently injecting malicious instructions into the System prompt; this change is stored in the model's Modelfile and is invisible to the client and undetectable by standard tools.65% similarVerifiedPrompt injection attacks can use carefully crafted user inputs to induce the model to ignore the constraints of the system prompt.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/963031API
curl https://kongchang.com/api/v1/knowledge/claims/963031MCP
get_claim(id=963031)