Attackers may infer or reconstruct original data by probing vector similarity APIs, especially with unrestricted access. This enables model inversion attacks, membership inference, and unauthorized data leakage from stored embeddings.
Adversaries may insert malicious or adversarial vectors into the index through ingestion endpoints, polluting the dataset and degrading search quality, or subtly steering results toward specific outcomes.
Attackers may infer sensitive information through metadata filters or by correlating embeddings across modalities (e.g., voice and face), bypassing surface-level access controls.
Attackers with unauthorized access or excessive permissions may tamper with or roll back index versions, potentially restoring poisoned data or breaking downstream integrations.
Poor validation of embedding formats or dimensions can cause service crashes or logic errors. This can result in denial of service or incorrect similarity results.
Approximate nearest neighbor (ANN) algorithms may yield non-deterministic or biased results. Adversaries may exploit these differences to evade detection or bias AI responses.