OpenAI Details GPT-Red: An Internal Automated Red-Teaming Model That Beat Human Red-Teamers 84% To 13% On Prompt Injection
84% vs 13%. OpenAI's GPT-Red just outperformed human red-teamers on prompt injection — and found a new attack class nobody saw coming.
Why it matters
OpenAI is automating adversarial testing at scale using self-play RL, achieving superhuman performance on prompt injection detection. This signals a major shift in how frontier labs approach safety validation and competitive red-teaming — but the model still struggles with multi-turn and image attacks, revealing gaps in current defenses.
The key facts
6 to knowGPT-Red achieved 84% vs 13% win rate against human red-teamers on prompt injection arena
Used self-play reinforcement learning against population of defender LLMs
Discovered novel 'Fake Chain-of-Thought' attack class
Reduced GPT-5.6 Sol failures 6x on OpenAI's hardest direct injection benchmark
Model acknowledges limitations: struggles with multi-turn and image-based attacks
Internal-only tool, not public release
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: OpenAI trained GPT-Red, an internal-only attacker model, using self-play reinforcement learning against a population of defender LLMs. It beat human red-teamers 84% to 13% on a replicated indirect prompt injection arena, found a novel "Fake Chain-of-Thought" attack class, and cut GPT-5.6 Sol's…