Kamus AI
Prompt injection
An attack in which text slips in orders to steer a model off course.
Prompt injection happens when text the model reads, such as an email, CV, web page or document, contains lines like "ignore your previous instructions". Models struggle to tell instructions from data. The defence is layered: mark outside text as data, limit the tools allowed, and require human confirmation for important actions.
Also called: Prompt injection, injection
Related terms
Appears in 15 lessons
- Projects, standing instructions and progress notes
- Connector risk and prompt injection
- Screening CVs with a rubric
- Prompts for job posts and screening
- From the need to the shortlist
- Invoices and tax documents
- Prompts for invoices, tax and approval
- AI features that read other people's text
- App security
- Session 5: security and review pass
- MCP from the inside, and how to choose a server
- Secrets out of the agent's reach
- Injection that arrives through data
- Layered fences
- Project: attack your own agent