OpenAI GPT-Red — Automated Adversarial Self-Play Red-Teamer for Prompt Injection

OpenAI disclosed (July 15, 2026) GPT-Red, an internal-only model trained via adversarial self-play to generate prompt-injection attacks against defender models at scale, with successful attacks fed back into training. OpenAI reports GPT-Red found successful attacks in 84% of held-out indirect-injection scenarios vs. 13% for human red-teamers, and that training against it made GPT-5.6 Sol 6x more resistant to prompt injection than OpenAI's best model four months prior; case studies include hijacking a live vending-machine agent and a Codex CLI data-exfiltration scenario.

Demonstrates a scalable AI-vs-AI red-teaming flywheel becoming standard practice at a top frontier lab specifically to harden production models against prompt injection — the top agentic-AI attack vector — with concrete before/after robustness numbers on a shipped model (GPT-5.6).

AI labs and enterprise agent builders should track this as a benchmark for prompt-injection robustness testing methodology; GPT-Red itself is not released publicly.