Data and Model Poisoning
Training data or a model is corrupted — before or during training — so it behaves maliciously, carries a backdoor, or is subtly biased.
What it means
Poisoning targets the data a model learns from, or the model’s weights, so the finished model behaves in a way the attacker chose — a hidden trigger, a bias, or a backdoor. It can also target any corpus a system later retrieves from.
Why it matters
Poisoning is hard to detect after the fact and affects every application downstream of the tainted model or dataset.
How we approach it
Low exposure by design. We don’t train or fine-tune models on your data, so there is no training pipeline for an attacker to poison here.
There is also no retrieval corpus or vector store to taint — the assistant reasons over your live application data at request time and then discards it, rather than learning from an index that persists.
The underlying models come from established foundation-model providers who carry the training-time responsibility for their own models.
Frequently asked
Do you fine-tune models on my data?
No. Your data is used only to answer the request in front of the assistant; it is never used to train or fine-tune a model.
Is there a knowledge base that could be poisoned?
No. There is no stored corpus or vector index. The assistant works with your live data through your APIs at request time, which removes the persistent-poisoning surface.