AI security guide
LLM Application Security Checklist
LLM security is not one vulnerability class. A useful review starts with the system boundary and works through the places where untrusted input, model behavior, tools, data, and application logic meet.
Updated September 14, 2026
01
Define the authorization boundary
Document the approved target, environment, accounts, tools, data, exclusions, and stop conditions before testing.
02
Trace untrusted input
Identify prompts, retrieved content, files, tool output, web content, and other data that can influence model behavior.
03
Minimize tool permissions
Give model-controlled workflows only the actions and data needed for the intended task, with explicit confirmation around consequential changes.
04
Test failure handling
Check what happens when the model is uncertain, a tool fails, output is malformed, or a request exceeds normal limits.
05
Keep evidence reviewable
Record the test version, request conditions, response evidence, confidence, and limitations so findings can be reproduced and challenged.
Next
Put this guidance into practice.
Evil AI's evaluator is designed around authorized, non-destructive checks with explicit coverage and uncertainty.
Explore automated AI red teaming → · More AI security guides →
Private beta · authorized applications only