What is Data and Model Poisoning?
Data poisoning manipulates pre-training, fine-tuning, or embedding data to introduce biased behavior, vulnerabilities, or backdoors. Model poisoning concerns compromised artifacts or weights. Both are integrity risks, but a malicious document retrieved at inference time is an application-input or knowledge-base-ingestion risk unless it changes learning data or the model artifact.
Data model poisoning risks should be assessed in the context of the system, data flows, identities, integrations, and decision consequences. A precise boundary helps owners evaluate the exposure without overstating what one control can achieve. See the authoritative source for the underlying reference.
What is Data and Model Poisoning used for?
Poisoning can degrade behavior or introduce targeted triggers that ordinary functional testing misses. Require provenance, versioning, integrity checks, supplier assessment, controlled ingestion, and use-case evaluation. Supply-chain security overlaps but is broader, covering dependencies and artifact handling. Do not imply that one dataset scan establishes model integrity.
Leaders should define accountable ownership, test relevant failure conditions, and retain evidence for decisions and change review. The response depends on the use case, authority, data sensitivity, architecture, and operational capacity rather than a universal checklist.
This assessment should be revisited when the system, data, model, connected services, permissions, or operating assumptions change.
