OpenAI unveiled GPT-Red, an in-house AI that hammers its own models to expose prompt-injection weaknesses before customers ever see them, according to SiliconANGLE. GPT-Red runs automated red-teaming, firing prompts at target models and refining its attacks through self-play reinforcement learning, where attacker and defender models train against each other. OpenAI said the system beats human red-teamers, landing 84% of its attempts against 13% for human testers, and cut direct prompt-injection failures to one-sixth of what its top model logged four months earlier. GPT-Red also broke autonomous agents in testing, hijacking a Vendy vending-machine agent and compromising command-line coding agents, yet OpenAI keeps it internal and unreleased, folding its findings into training.
