Vector databases have become the retrieval backbone for modern AI applications, but their security has received little attention. A new arXiv paper introduces the Black-Hole Attack, a poisoning attack that works by injecting a small number of malicious vectors into a vector database.
The attack exploits defects in embedding space, causing queries to be pulled toward attacker-chosen content. Because the attack needs only a small number of injected vectors, it may be difficult to detect in practice.
The paper serves as a warning that embeddings are not neutral infrastructure: database operators should not assume the vectors they store are trustworthy. Since this is a single arXiv announcement, there are no independent results to compare, and the findings have not yet been verified by other research groups.