OpenAI’s GPT-Red Cuts Prompt Injection Success Rate from 90% to Under 23%

OpenAI’s GPT-Red Cuts Prompt Injection Success Rate from 90% to Under 23%

A single number summarizes this week’s most significant AI security development: 23%. That’s the new success rate for prompt injection attacks on GPT-5.6, down from 90% to under 23%.

A single number summarizes this week’s most significant AI security development: 23%. That’s the new success rate for prompt injection attacks on GPT-5.6, down from 90% to under 23%.

OpenAI’s GPT-Red automated red-teaming model reduces prompt injection attacks on GPT-5.6 to under 23%. Plus: Veo 3 goes public, Gemini delays, and a new global AI body.

As a result, the company has made significant strides in AI security, reducing the risk of prompt injection attacks. This week, OpenAI announced that it has successfully reduced the success rate of prompt injection attacks on its latest model, GPT-5.6, to under 23%. This is a significant improvement from the previous success rate of 90%.

OpenAI’s GPT-Red model has been designed to identify and mitigate potential vulnerabilities in AI systems. The model uses advanced techniques to simulate various attack scenarios and assess the effectiveness of the AI’s defenses.

In addition to the improvements in prompt injection security, OpenAI has also made strides in other areas of AI safety. The company has been actively working on developing new frameworks and guidelines for responsible AI deployment.

As AI technology continues to evolve, the need for robust security measures becomes increasingly important. OpenAI’s efforts to enhance the security of its models are a step in the right direction, ensuring that AI systems can be deployed safely and responsibly.