Skip to content
ML Academy

Introduced on day 47

Prompt injection

Text placed in a model's input, directly by a user or indirectly through fetched content, that tries to make the model follow the attacker's instructions instead of the application's.

In detail

Direct injection is typed by the user; indirect injection sits in content the application reads, such as an email, a web page or a retrieved passage. Retrieval and fine-tuning do not remove the risk. Effective controls limit what a compromised output can do: narrow tools, validated arguments, least privilege, approval and downstream authorisation. Day 53 treats the threat model in depth.