SECAI Core · phase 11 of 15
Model poisoning
A security attack where harmful data or code is intentionally added during the model's training or fine-tuning process
The Explain card
- Plain English
- Model poisoning is intentionally inserting harmful data or code during a model's training or fine-tuning so the finished model behaves the way the attacker wants.
- Example
- A contractor fine-tuning a customer-service model slips in examples that teach it to reveal account details whenever a user types a particular passphrase.
- Why it matters
- Training pipelines involve many hands and many steps. Defenders secure the fine-tuning process, review datasets and training code, and evaluate models against red-team suites before deployment.
- Hook
- Tampering with the recipe, not the meal.
Where it sits in the deck
Phase 11: Threat Landscape: Attack Vectors and Adversarial Techniques
With defences named, learn what they must defend against — the full catalogue of attack techniques targeting AI systems, their inputs, outputs, training pipelines, and supply chains.
