Prompt injection happens when malicious or untrusted content tells an AI system to ignore its instructions, reveal secrets, misuse tools, or take actions the user did not intend.
In practice
It often appears in web pages, emails, documents, tickets, or database content that an AI agent reads. The attacker hides instructions inside data the model treats as context.
What to watch
Do not rely only on prompts to stop prompt injection. Use tool permissions, data separation, allowlists, confirmations, and output validation.