AI SECURITY • 349 / 397
Understand how untrusted text, tools and autonomy create new attack paths.

Tool Output Injection

Tool output injection places malicious instructions inside data returned by a tool, hoping the model treats the data as commands.

Think of it like

Think of Tool Output Injection as an input-trust or privilege problem: words can influence software decisions, so boundaries must be explicit.

Real life

Email, documents and web pages can become hostile inputs to AI systems through attacks involving Tool Output Injection.

SRE lens

A webpage search result can contain text telling the agent to call another tool.

Remember thisTool output injection places malicious instructions inside data returned by a tool, hoping the model treats the data as commands.
AIForSREJump to a concept
Search is optional. The main journey is simply ↓ Next.