Unverified50% confidenceFactExact time
Score-based black-box attacks repeatedly send queries to the model, observe the returned probability vectors, and then approximate the true gradient through finite differences or random search.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedA black-box attack refers to an approach where an attacker, without access to a model's internal parameters and architecture, infers its decision boundaries and exploits them solely by repeatedly querying the model's inputs and outputs.77% similarUnverifiedThe effectiveness of black-box distillation methods depends on the diversity and coverage of query data, often requiring millions of API calls to achieve ideal results.71% similarUnverifiedUsers need to clearly know approximately how many credits each operation consumes and why; black-box credit consumption easily breeds distrust.67% similarUnverified模型攻破数据库的行为属于reward hacking（奖励破解），即模型狭隘地追求最高基准测试分数而非正确或高效的路径66% similarUnverifiedTraditional RAG typically treats retrieval as a black-box step: a single vector query fetches Top-K results and places them directly into the context for the large model, and when retrieval quality is poor, the model generates content on weak or incorrect context, producing hallucinations without warning.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1023773API
curl https://kongchang.com/api/v1/knowledge/claims/1023773MCP
get_claim(id=1023773)