Vector embedding security gap exposes enterprise AI pipelines

VectorSmuggle is a research framework demonstrating that sensitive corporate content converted into vector embeddings for internal AI assistants can be used as a covert exfiltration channel: attackers with pipeline access can encode payloads into embeddings (via noise, scaling, rotation, model-splitting, etc.) that still function for legitimate search but carry hidden data. Some perturbation methods are detectable by anomaly detectors, but rotation-preserving transforms evade detection and can carry large payloads; the payloads survive popular vector stores. The project also proposes VectorPin, a cryptographic signing scheme to detect tampering of embeddings.