GPT-Red对抗训练提升GPT-5.6提示注入鲁棒性

Highly recommended read.

This is another great high ROI application of self-improving AI models.

GPT‑Red adversarially trained GPT‑5.6, which made it much more robust to prompt injections.