Unverified50% confidenceFactExact time
The GCG (Greedy Coordinate Gradient) attack was proposed by a Carnegie Mellon University team in 2023; it uses gradient optimization to generate text suffixes that bypass safety restrictions, and the attack can be transferred from open-source models to closed-source commercial models.
1
Sources
50%
Confidence
Long-term
Relevance
8/24/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedIndirect prompt injection conceals attack instructions within external documents, web pages, or database records retrieved by the model, and is activated when a RAG system introduces them into the context59% similarUnverifiedRAG架构引入了间接提示注入攻击面：攻击者只需在RAG检索的数据源中植入恶意指令即可操纵模型58% similarUnverified对抗样本（Adversarial Examples）由Goodfellow等人于2014年首次系统描述，经典攻击算法包括FGSM和PGD58% similarUnverifiedClaude Code discovered deep-level resource leak risks such as ClassLoader isolation potentially preventing GC and unbounded growth of index metric data with no automatic cleanup.57% similarUnverifiedClassic GNN variants include GCN (Graph Convolutional Network), GraphSAGE (Sample and Aggregate), and GAT (Graph Attention Network), with their main differences lying in how the aggregation function is designed.57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/796914API
curl https://kongchang.com/api/v1/knowledge/claims/796914MCP
get_claim(id=796914)