Unverified50% confidenceDecision RuleExact time
Any text originating from external, untrusted sources should be treated as a potential attack vector before entering a large language model.
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
Unverified自然语言作为攻击载体使得攻击样本几乎无法被签名检测，每次攻击可用不同措辞表达相同恶意意图73% similarUnverifiedPrompt Injection is one of the most closely watched attack surfaces in the current large language model security field.72% similarUnverified强对抗性攻击如使用另一个模型完整改写或翻译往返仍可能破坏水印信号，鲁棒性与文本质量之间存在权衡71% similarVerifiedLarge language models are essentially probabilistic text generation systems, lacking the ability to verify external systems in real time, and produce hallucinations when input documents are ambiguous or incomplete.69% similarUnverified提示注入（Prompt Injection）的核心原理在于大语言模型无法可靠地区分用户的真实指令和内容中夹带的恶意指令68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/910263API
curl https://kongchang.com/api/v1/knowledge/claims/910263MCP
get_claim(id=910263)