During a security review, a team discovers that their AI model training pipeline pulls data from a shared internal data lake with no integrity verification on ingested files. Which attack vector does this create and what control mitigates it?

Question

Accepted Answer

A. Data poisoning risk — mitigated by implementing cryptographic integrity checks and access controls on training data sources. Without integrity verification on ingested training data, a threat actor with write access to the data lake can inject malicious samples — a data poisoning attack. Cryptographic hashing or digital signatures verify data has not been tampered with, and access controls limit who can modify it. The SecAI+ objectives cover securing data pipelines as part of protecting AI deployment environments. Model inversion is an inference-time privacy attack unrelated to training data ingestion.

During a security review, a team discovers that their AI model training pipeline pulls data from a shared internal data lake with no integrity verification on ingested files. Which attack vector does this create and what control mitigates it?

Related Questions

Discussion