The developers of GPT‑5.6 increased the model's resilience by training against GPT‑Red: they replayed GPT‑Red's strongest attacks, which the model had not seen during learning. GPT‑5.6 Sol proved effective, producing six times fewer errors in prompt injection tests compared with the best deployed model from four months earlier, an important advance in model safety.
Training against GPT‑Red: GPT‑5.6 Sol became significantly more resilient
The developers of GPT‑5.6 increased the model's resilience by training against GPT‑Red: they replayed GPT‑Red's strongest attacks, which the model had not seen during learning.
