dailyai.report

The internal GPT-Red model successfully attacked test scenarios 84 percent of the time. Human red teamers managed only 13 percent. This self-play training identifies vulnerabilities far faster than manual testing. OpenAI now uses these automated findings to harden the safety layers of its next-generation models before public release.