AI models and agents

Prompt injection

Also known as: instruction injection

An attempt to change AI behavior through instructions hidden in untrusted data.

A practical example

A document contains instructions to ignore the user's task.

What should I know?

Source content should not grant action permissions by itself.

AI models and agents in practice

A model processes information; the application defines tools and permissions. Reliable answers require good sources and verification. Actions involving accounts, content, and orders must stay within the granted scope.

Related Semly features

See where this concept applies when working with Semly.

Semly MCPConnect Semly data and tools to your AI assistant. Analyze reports and work with Leon in a conversation.Websites for AI agentsShare structured company knowledge and inquiry forms through an agent channel.

Related terms

AI agentA system using a model and tools to pursue a goal within available capabilities.AuthorizationDetermining which resources and operations an identified user or application can access.Tool callingA mechanism for a model to propose using an available function, which the application executes.

Documentation and further reading

Definition prepared by Semly. Standards and behavior are described by their creators:

Find out if AI ChatGPTGeminiClaudeGoogle AI OverviewsGoogle AI Mode can see you

Enter your website address to get your AI visibility report.