Model poisoning
Model poisoning is the deliberate corruption of an AI system by feeding it manipulated data — during training or retrieval — so that it learns false information or behaves the way an attacker intends.
How to recognise
It is the broader category that techniques like "LLM grooming" belong to. Poisoning at the retrieval layer — what a model pulls from the live web — can often be mitigated by filtering bad sources. Poisoning at the training layer is far more serious, because false information absorbed during training is baked in and cannot simply be removed afterward.
For organisations building or relying on AI, the defence is data-provenance discipline: vetting where training and retrieval data comes from, and excluding known manipulation sources before they are ingested.
Related warrier.ai coverage
Disinformation · Grooming the Machines: How the Pravda Network Targets the AI You Ask for the Truth