Unverified50% confidenceFactExact time
The GCG attack (Greedy Coordinate Gradient Attack) is an adversarial prompt attack method that, through gradient-guided greedy search, progressively replaces tokens in the input sequence to maximize the probability that the model generates the target harmful content.
1
Sources
50%
Confidence
Long-term
Relevance
10/28/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe 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.79% 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69% similarUnverifiedFor random-search-type attacks, GCD induces systematic accept/reject errors, that is, the attacker frequently makes wrong judgments when determining whether a perturbation is effective.67% similarUnverifiedEcho-like attacks are a typical weakness of RAG architectures: attackers disguise malicious instructions as retrieved result fragments, and the model mistakes them for legitimate instructions and performs operations such as deletion and data exfiltration.62% 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60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1105874API
curl https://kongchang.com/api/v1/knowledge/claims/1105874MCP
get_claim(id=1105874)