TLDR Tech / InfoSec
GPT‑Red: Unlocking Self-Improvement for Robustness
OpenAI trains GPT‑Red to run automated prompt‑injection attacks against internal defender models and live agent setups, such as a vending machine controller and a Codex CLI assistant, then uses those attacks to harden GPT‑5.x models against data exfiltration and tool misuse. GPT‑5.6 now shows sharply lower failure rates on direct and indirect injections while keeping normal capabilities intact.
Read the full story