Skip to content
Get in touch

Needs review · content version 2

Review Model Poisoning

Your draft is bound to this exact source hash. If the term changes, the Review Desk will use a new storage key instead of silently attaching old decisions.

Definition

Model poisoning is a deliberate attack that corrupts training data, learning processes, model updates, or model artifacts to alter behavior or introduce hidden weaknesses.
Decision for Definition

Plain-Language Explanation

The attacker interferes before or during model development rather than only at the moment a user sends a request. The effect may be broad degradation or a targeted behavior triggered under specific conditions.
Decision for Plain-Language Explanation

Why It Matters

A compromised model can pass ordinary tests while retaining a hidden failure path. Supply-chain controls, provenance, access restrictions, and targeted evaluation matter.
Decision for Why It Matters

Example

An attacker inserts manipulated examples into a training pipeline so a particular visual pattern causes the model to assign a chosen label.
Decision for Example

Common Misconception

Model poisoning is different from prompt injection. Poisoning alters training or model integrity; injection attacks a deployed system through processed instructions.
Decision for Common Misconception

Remember It This Way

Poisoning changes what the model learned or carries.
Decision for Remember It This Way

Related Terms

* [[training-data|Training Data]] * [[data-provenance|Data Provenance]] * [[capability-provenance|Capability Provenance]] * [[prompt-injection|Prompt Injection]] * [[adversarial-attack|Adversarial Attack]]
Decision for Related Terms

Sources

* [NIST AI 100-2 E2025, Adversarial Machine Learning](https://doi.org/10.6028/NIST.AI.100-2e2025) — authoritative taxonomy of poisoning attacks and mitigations.
Decision for Sources

Learn More

* Use [[data-provenance|Data Provenance]] to examine the training supply chain.
Decision for Learn More
Return to term

Not saved yet.