liveBook · Manning

Poisoning attack

As explained in AI Governance

Model poisoning is the deliberate or accidental introduction of harmful, misleading, or biased data or updates that alter a model's behavior.

A poisoning attack deliberately or accidentally introduces harmful, misleading, or biased data or updates that alter a model’s behavior. It is a model-weight risk: the vulnerability can reside in what a model has learned rather than in code that runs when its file is loaded. section 4.4.1

Poisoning is identified as one path to weight-based risk, distinct from safety degradation through fine-tuning. The attack’s mechanisms and mitigations are not detailed here. section 4.4.1

How AI Governance covers it

How deep it goes
Defines poisoning and distinguishes it from fine-tuning degradation, but stops short of attack mechanics or mitigations.
What you can do after
Understand poisoning as a behavior-altering risk embedded in a model rather than external code executed during loading.
How it connects
Connects poisoning to model weights and contrasts it with fine-tuning and safety degradation.

Also explained in

Compare how 4 books explain poisoning attack →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.