OpenAI Deploys GPT‑Red to Auto‑Find Prompt‑Injection Flaws and Harden GPT‑5.6
OpenAI’s new GPT‑Red runs an AI attacker against its own models, brute‑forcing thousands of prompt‑injection variants in seconds. The discovered exploits are injected into the training loop of GPT‑5.6, slashing its failure rate on a hard benchmark by six times. This automated red‑team adds scalable safety to agentic systems.