OpenAI Deploys GPT-Red Automated Red-Teaming Model to Harden GPT-5.6 Defenses
Tags AI · Security · OSS
OpenAI revealed GPT-Red, an internal LLM trained via self-play to automate red-teaming of its models. In replication of a 2025 human red-team exercise against GPT-5, GPT-Red discovered more effective attacks than human testers, including a novel 'fake chain of thought' prompt injection technique. Training GPT-5.6 against GPT-Red reduced attack success from over 90% (GPT-5) to under 23%. GPT-Red will not be released publicly.
Technical significance
Automated red-teaming via self-play represents a scalable approach to AI safety evaluation that can keep pace with model capability growth. The 'fake chain of thought' attack — injecting fabricated reasoning steps into a model's scratchpad — reveals a new vulnerability class in chain-of-thought architectures. The 90% to 23% attack success reduction demonstrates measurable hardening, though GPT-Red's weakness in multi-turn conversational attacks and image-based injections indicates coverage gaps.