Home›Crypto›Regulation›OpenAI rolls out GPT-Red to harden GPT-5.6 against pro…

OpenAI rolls out GPT-Red to harden GPT-5.6 against prompt injection

OpenAI said GPT-Red succeeded in 84.0% of internal evaluation scenarios, versus 13.0% for human red teamers, and helped train GPT-5.6.

Reported · Written up by our desk · Regulation

OpenAI introduced GPT-Red, an automated red-teaming system designed to identify security vulnerabilities in its language models before release, and said it was used to train GPT-5.6 to reduce failures on prompt injection tests.

According to Decrypt, OpenAI said GPT-Red uses adversarial self-play, generating progressively stronger prompt injection attacks while defender models learn to resist them, and that the identified attacks were incorporated into GPT-5.6 training.

OpenAI reported GPT-Red succeeded in 84.0% of internal evaluation scenarios compared with 13.0% for human red teamers in the same tests, describing the approach as a way to scale red-teaming beyond what it can accomplish with human teams alone.

As an example, OpenAI said GPT-Red manipulated an autonomous vending machine agent to lower prices, order discounted inventory, and cancel another customer’s order before the vulnerabilities were disclosed and addressed.

More like this

Sources