🛡️ Security · ⏱️ 2 min

What is Data poisoning?

Simply put, like I'm 5 years oldIt is like someone slipping a few drops of a bad ingredient into a big pot of soup while it cooks. You cannot see it, but later the soup makes people sick in a way nobody expected.

📖 The slightly longer version

An AI model learns from a huge pile of examples, called training data. Much of it comes from the open web. Data poisoning means someone sneaks bad examples into that pile on purpose. The model learns from them like it learns from everything else.

The worst kind plants a backdoor. A backdoor is a hidden behavior that stays quiet until a secret trigger appears, like a special word or phrase. Most of the time the model acts normal and passes its tests. Then the trigger shows up, and it does the bad thing. Today's paper, 'Agentic backdoors via pretraining poisoning', looks at this during pretraining, the first big learning phase. It also looks at agents, AI that takes actions for you, not only writes text.

How a poisoned model is made

Bad data plantedAn attacker puts a few crafted examples where training data is collected.
Model trainsThe model learns from the whole pile, bad examples included.
Looks normalThe model passes its tests and behaves well most of the time.
Trigger appearsA secret word or situation shows up in a request.
Hidden behaviorThe model, or the agent using it, does what the attacker wanted.

Why it matters

Today we also read about agents moving onto phones and tools with access to our apps. If the model inside an agent carries a hidden backdoor, it could act fine for weeks and then do something harmful when it sees a trigger. You cannot spot this from the outside, so it matters who builds the model, where its data comes from, and who checks it.

⚠️ What to watch out for

This is one research paper, and our sources give few details, so we do not know how easy the attack is in real life or how big the risk is today. Researchers show that attacks are possible; that does not mean they are already happening widely. Still, nobody has a proven way to find every hidden backdoor, and labs rarely share exactly what data went into their models.

📰 Where it showed up

This notion came up in AI starts editing AI, while the money and the attacks keep growing. More lessons: Learn.

🎯 Keep learning

🛡️ Security · 2 min What is Supply Chain Attack? Instead of breaking into your house, a thief hides something bad inside the milk the grocery store delivers to you. You trust the store, so you open the door and take it in. Learn it →

🛡️ Security · 2 min What is Coordinated Disclosure? Imagine you find a hole in your neighbour's fence. You tell them quietly first, so they can fix it before you tell everybody, and before anyone sneaks through. Learn it →

🛡️ Security · 2 min What is Exploit? A bug is a broken lock on a door. An exploit is someone actually opening the door with a trick and walking in. Learn it →

Sources

arXiv

Written on 2026-11-03. This explanation was accurate on that day. AI changes fast, so check the sources and the date before you rely on it.

Get The AI Daily Grind by email, every morning, in simple words.