Security basics for shipping AI features
LLM features add new attack surface. A handful of habits handle the most common risks before they reach users.
An AI feature takes untrusted input and often gives the model tools that can act. That combination is the whole security story: treat model input as untrusted, and constrain what the model is allowed to do.
Never put secrets or another user's data into a prompt the model might echo back. Validate and constrain tool calls the model can make, and apply the same authorization checks you would for any other code path. A model asking to read a file is not permission to read any file.
Guard against prompt injection: content the model reads can contain instructions aimed at it. Do not let retrieved documents or user text silently change what the system does. Log model actions so you can review what happened.
Key points
- ·Treat model input as untrusted; constrain model actions.
- ·Never expose secrets or other users' data in a prompt.
- ·Assume retrieved content may carry injected instructions.
Sources
Replies
Sign in to reply.
No replies yet. Be the first to add something useful.