Slide 6 of 27
Part 1 · What Is It?Slide 6
Slide 6 · Who Does This
Who poisons a model — and why.
It isn't always a shadowy hacker. Sometimes it's an artist. Sometimes it's a crowd.
🕵️
The Supply-Chain Attacker
MALICIOUS
Uploads a tampered model or dataset to a public hub, hoping you'll download and trust it.
🤬
The Coordinated Crowd
MALICIOUS
Floods a system that learns from user input with toxic content — like the trolls who broke Tay.
🎨
The Defensive Artist
PROTECTIVE
Poisons their own work (e.g. Nightshade) to stop AI from training on it without consent.
🤷
The Careless Contributor
UNINTENTIONAL
Feeds unverified or biased data into a pipeline with no malice — poisoning by accident.

That last one matters: not all poisoning is an attack. Unvetted data and sloppy pipelines poison models every day without anyone meaning to.

← BackNext → Poisoning vs injection