Can You Trust the Vectors in Your Vector Database? Black-Hole Attack from Embedding Space Defects
The Black-Hole Attack is a poisoning method where malicious vectors are placed near the geometric center of a vector database. This exploits 'centrality-driven hubness,' a structural property of high-dimensional space where central points become disproportionately influential. The attack shows that vectors cannot be trusted blindly, as geometric flaws make retrieval vulnerable to widespread poisoning.