Model Poisoning: Teaching an AI the Wrong Lessons on Purpose
If the data that trains or feeds a model can be tampered with, so can its behaviour.
2 min readAdvanced Sep 10, 2026
Explain Like I'm Not a Hacker
Slip a few wrong answers into an answer key, and everyone who studies from it learns them.
The 30-second explanation
If you teach a student from a book with a few pages secretly rewritten, they will confidently repeat the wrong things. Poisoning does that to AI.
How it works
- 1
1. Find a source
Identify data the model relies on.
- 2
2. Tamper
Insert misleading or malicious content.
- 3
3. Learn or retrieve
The system uses the tampered data.
- 4
4. Behave
It produces the attacker's desired output.
Models learn from large amounts of data, and increasingly retrieve extra information at answer time. An attacker who can influence those sources, for example by editing public pages or contributing malicious examples, can shift behaviour. The tampering may be subtle and only show up for a specific input. Defenders track where data comes from, check what goes in, test models for unexpected behaviour, and restrict who can change datasets.
Real-world example
A system that answers questions from a shared knowledge base is fed a document with subtly wrong instructions. Later, employees following the assistant's confident answer perform a step that weakens a security setting.
How to spot it
Unexplained accuracy changes
A sudden shift in behaviour after a data update.
Unknown data sources
Content with no clear origin or owner.
Odd behaviour on specific inputs
A trigger phrase that changes results.
Uncontrolled edit access
Many people able to change the data the AI trusts.
What to do
- 1Record where every dataset comes from, and restrict who can change it.
- 2Test models against a fixed set of checks after each data or model update.
- 3Treat externally sourced content as untrusted until it has been reviewed.
Stay curious. Stay safer.
This is one piece of a bigger picture. Explore more real-world examples, concepts and tips to build your cybersecurity awareness.