Guardrails for untrusted retrieved text, tool output, and user-supplied instructions.
Publié Sep 13, 2026
MIT · Published Sep 13, 2026