AI Data Security

AI systems depend on data far more deeply than conventional applications. They do not merely store and retrieve it — they encode it into parameters, compress it into vectors and concatenate it into context. Every one of those transformations adds a layer of risk that conventional data-security controls were never designed to cover.

Three stages, three kinds of risk

Training is about provenance and integrity. Where the corpus came from, whether it was poisoned, whether it carries hidden trigger patterns — the answers stay in the model weights, and they are very hard to remove afterwards.

Retrieval and storage is about what vectorisation breaks. Once text becomes an embedding, the original access-control and redaction policies often stop applying: a vector can be inverted back to approximate meaning, and badly designed retrieval permissions make unauthorised access extremely quiet.

Inference is about runtime leakage. Context left over from earlier turns, the persistence policy for logs and caches, and the strength of tenant isolation together decide whether one ordinary request can surface somebody else’s data.

Planned coverage

• Training-corpus security: data poisoning and backdoor insertion techniques, plus provenance, licensing and compliance risk.
• Vector stores and retrieval augmentation: vector inversion attacks, unauthorised retrieval, embedding leakage, and out-of-scope splicing of retrieved chunks.
• Leakage on the inference path: context carry-over, log and cache policy, and the common failure points in multi-tenant isolation.
• Data residency and compliance: cross-border transfer, data minimisation, and what redaction and pseudonymisation actually look like in AI systems.
• Sensitive-data detection and masking: the engineering practice of identifying and handling sensitive content before it reaches the model.

Articles in this area are still being researched and verified. They will be published here as they are finished.