Adversarial Fine-Tuning for Prompt Injection Defense in GPT-3

In Sep 2025 researchers revisited a 2022 study and found fine‑tuning cut prompt‑injection success from ~31% to near zero on GPT‑3 Ada, Babbage and Curie, while Davinci stayed vulnerable. getnews.me/adversarial-fine-tuning-... #gpt3 #adversarialfinetuning