Data poisoning attacks insert crafted examples into training data to cause targeted misbehavior or general degradation. Backdoor attacks introduce triggers that produce specific outputs when encountered.
Risk is highest where training data is collected from open sources.
Alternative Names:
Training Data Poisoning, Poisoning Attack
Why it Matters?
For legal buyers this is primarily a vendor diligence question rather than a direct operational risk, since firms rarely train models. The relevant inquiries are whether the vendor controls its training data sources and whether it fine-tunes on customer-supplied content, which creates an ingestion path. Retrieval-based systems face a related but distinct risk, since poisoned content in a document corpus can influence outputs without touching model weights.
Frequently Confused with
Related terms
Frequently asked questions
Is data poisoning a direct risk for law firms?
How does it affect retrieval systems?
