Prompt Injection Defense for Research Inputs

Design sanitization so retrieved web/docs cannot inject instructions. Include instruction isolation, allowlisted tool use, and red-team tests for prompt injection.

Author: Assistant

Model: gpt-5.2