OpenAI's GPT-Red: AI that attacks itself to get safer
OpenAI just announced GPT-Red, an internal AI system that attacks its own models to find prompt injection failures before they reach users.
Transcript
OpenAI just announced GPT-Red, an internal system that attacks its own models to find safety failures.
GPT-Red is an internal OpenAI tool, not a public product or API. It generates attacks, then scores the failures, then hardens the model against them, then repeats the loop.
On OpenAI's own internal prompt injection arena, GPT-Red succeeded eighty four percent of the time. Human red teamers reached only thirteen percent.
After GPT-Red training, GPT 5.6 Sol had six times fewer failures than the prior best model. Direct injection failures dropped to zero point zero five percent.
Inboxsmith uses tested AI to answer your calls so you never miss a lead. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- GPT-Red is an internal OpenAI system that uses self-play to find prompt injection failures.(GPT-Red: OpenAI's Internal AI Red Teamer Explained)
- GPT-Red reached 84 percent attack success versus 13 percent for human red teamers on OpenAI's internal indirect prompt injection arena.(GPT-Red: OpenAI's Internal AI Red Teamer Explained)
- GPT 5.6 Sol had six times fewer failures than the prior best production model, and direct injection failures dropped to 0.05 percent.(GPT-Red: OpenAI's Internal AI Red Teamer Explained)
- A green test is not a safety certificate because attackers keep changing while static tests do not.(GPT-Red and the end of one-off AI security testing)