Safety and MLOps
Data poisoning
Data poisoning manipulates training or retrieval data to change system behavior.
How it works
An attacker can insert mislabeled examples, backdoors, or malicious documents. Provenance, filtering, and anomaly detection reduce exposure.
A concrete example
Illustrative: 200 poisoned records are inserted into a 100,000-record training set.
Do not confuse it with Prompt injection
poisoning changes stored data; prompt injection influences a particular model interaction.