What the term really means
Deliberate corruption of training, retrieval or input data intended to influence model output.
How it works in practice
Poisoned data can introduce a hidden pattern into training, a search index or RAG documents and alter answers only under selected conditions.
The decision to make before implementation
Protection requires data provenance, change control, source separation and the ability to roll back a defective batch.
How to verify that it works
Watch for sudden output shifts, unusual retrieved documents, quality anomalies and gaps between a trusted test set and production. Compare results with an agreed baseline and review routine cases, difficult exceptions and human hand-offs separately. A practical Data Poisoning test should have an owner, a review date and a recorded example of an outcome the team will not accept.
The PAR HOUSE Agency approach
We approach Data Poisoning from the workflow rather than a tool demonstration. Protection requires data provenance, change control, source separation and the ability to roll back a defective batch. We then build a small measurable scope, record assumptions and expand only after quality review.
