OpenAI is now using AI to attack its own AI, and it's working better than humans ever did
84%. That's how often OpenAI's GPT-Red finds vulnerabilities in its own models—6x better than human red teamers. The future of AI safety isn't human oversight. It's AI attacking AI.
Why it matters
OpenAI has shifted from human-led red teaming to AI-powered adversarial testing, dramatically improving vulnerability detection rates. This signals a fundamental change in how frontier labs approach model safety and hardening—moving from manual processes to automated, scalable attack discovery that feeds directly into production model improvements.
The key facts
5 to knowGPT-Red achieves 84% success rate finding attacks in test scenarios
Human red teamers achieve 13% success rate by comparison
Self-play training methodology used for vulnerability discovery
Results integrated into GPT-5.6 Sol hardening pipeline
Represents 6.5x improvement over human red teaming baseline
Go to the source
The Decoderthe-decoder.com
Publisher excerpt: OpenAI's internal GPT-Red model finds successful attacks in 84 percent of test scenarios through self-play training. Human red teamers manage just 13 percent. The results feed directly into hardening models like GPT-5.6 Sol.