Holon.liveHOLON OF HOLONS

What is Prompt Injection? Content and instructions

English · Compare Chinese

Prompt Injection occurs when input or external content tries to redirect intended model behavior. For agents using webpages, files and tools, it can affect actions as well as answers.

Commands inside material

A webpage can contain imperative text without gaining authority over the user's task. Research systems need to understand what a page describes, not treat every command as authorization.

Data and authority

Tasks come from authorized parties. Retrieved material, documents and tool outputs are content to assess. A claim to be an administrator notice does not establish authority.

Evaluating defenses

OWASP discusses prompt injection as an important risk. Measures need assessment for actual systems and tasks; one added prompt cannot guarantee a complete solution.

Does a trusted source eliminate the risk? No. Consider the source, the content and how the system uses it.

Continue with Agent identity, delegation and authorization to compare related ideas.

For a related project perspective, explore LingoAI’s agents work.

Sources and further reading