Unverified50% confidenceOpinionExact time
In jailbreak attack scenarios, scoring an early prefix alone may not accurately predict its attack effectiveness when combined with a suffix, so pairwise evaluation is more reliable than evaluating prefixes individually.
1
Sources
50%
Confidence
Long-term
Relevance
10/20/2026
First Seen
Sources
Related Entities
Related Claims
Unverified针对特定触发词的后门攻击可让模型在大多数情况下表现正常，仅在遇到特定输入时才输出预设错误信息，使常规评测基准难以捕捉72% similarUnverifiedReference model-free attacks achieve, in most scenarios, results statistically comparable to those of attacks that rely on adversarial knowledge.67% similarUnverifiedPoisoning attacks designed for exact retrieval fail on the approximate HNSW indexes used in production, but the topology attack proposed in "Poisoning the Neighborhood" can exploit the structural bias of approximate indexes themselves and is harder for traditional content-filtering defenses to identify.65% similarUnverified预标注会引入锚定效应，标注员倾向于接受而非推翻已有预标注结果，即使存在微小错误，低质量预标注会系统性地将判断拉向错误方向64% similarUnverifiedIndirect prompt injection specifically refers to attackers carrying out attacks through external content that the model actively retrieves, making it more dangerous.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1074850API
curl https://kongchang.com/api/v1/knowledge/claims/1074850MCP
get_claim(id=1074850)