Module 2 of 5
How attacks work
The four families of attack on AI systems at the level of mechanism: injection, exfiltration, poisoning and supply chain, and abuse of the model as a resource.
You cannot defend against something you cannot picture. This module explains the four families of attack on AI systems at the level of mechanism: what the attacker controls, what they write, and why it works. Every defense in the later modules then has a clear target.
By the end of it you should be able to take any AI feature and say how text could hijack it, how data would leave once it was hijacked, what it was built from that someone else could have tampered with, and what it would cost if an attacker simply used it hard.
The demonstrations are small Python programs that use only the standard library and run offline. Attacks and the first defenses against them run against simulated models and services inside each program, so you can see the shape of an attack without targeting anything real.
Lessons
- 2.aPrompt injectionHow does text hijack a model, directly and indirectly?16 min
- 2.bData exfiltration and leakageHow does data get out once a model has been steered?14 min
- 2.cPoisoning and supply chainHow are models, data, tools, and memory corrupted before you ever run them?15 min
- 2.dAbuse and resource attacksHow are jailbreaks, extraction, and denial of wallet carried out?14 min