OpenAI has built a large language model designed to act as a super-hacker, known as GPT-Red. The company uses this system as an automated red teaming tool to test and improve the security of its other artificial intelligence models.
GPT-Red serves as a sparring partner for OpenAI’s existing models. By simulating cyberattacks, the system helps boost defenses against potential security threats. OpenAI states that it uses GPT-Red specifically to make its models safer.
Impact on GPT-5.6
The development of GPT-Red coincides with the release of GPT-5.6, the latest version of OpenAI’s flagship large language model, which was released last week. According to OpenAI, training GPT-5.6 against GPT-Red resulted in the company's most robust release to date.
During these tests, GPT-Red uncovered specific vulnerabilities. These findings were used to make GPT-5.6 more resistant to prompt injection attacks, a common method for manipulating AI systems.
It is indicated that GPT-Red uses self-play mechanisms to improve AI safety, alignment, and robustness against prompt injections. This claim remains unverified by independent sources.
