Prompt Injection

Prompt injection is an attack in which malicious instructions hidden in input data trick a language model into ignoring its original instructions and acting against intent.

Because models treat retrieved content and user input as text to follow, an attacker can embed commands in a web page, document, or message that the agent then obeys.

It is a leading security risk for tool-using agents, since a successful injection can cause an agent to leak data or take unauthorized actions; guardrails and input isolation are common defenses.