Unverified50% confidenceFactExact time
When a model receives poisoned input (prompt injection or indirect injection), it may be induced to perform actions the user never authorized.
1
Sources
50%
Confidence
Long-term
Relevance
9/29/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedPrompt injection is divided into two categories: direct injection (the user directly manipulates model behavior) and indirect injection (contaminating the context through external data sources).69% similarVerified提示注入是一种针对大语言模型的攻击手法，攻击者通过在模型输入数据中嵌入恶意指令劫持模型的正常行为67% similarUnverifiedBackdoor attacks plant triggers in training data, causing the model to behave normally on regular inputs but produce preset malicious outputs when specific trigger words are encountered.66% similarUnverified某些越狱攻击的有效性来自于诱导模型进入特定吸引子状态，使输出在该框架内自我强化64% similarUnverifiedAgent behavior is influenced by probabilistic model outputs, and the same user request may follow different execution paths due to differences in temperature parameters, context windows, or model versions.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/958267API
curl https://kongchang.com/api/v1/knowledge/claims/958267MCP
get_claim(id=958267)