OpenAI trains AI to attack other artificial intelligences
OpenAI surprised everyone by launching GPT-Red, an artificial intelligence designed to attack other AIs. It seems contradictory, but the idea is to identify vulnerabilities in language models and reinforce their security. Imagine a digital hacker, but with the goal of protecting, not destroying. However, this tool will not be made available to the public. OpenAI intends to use it exclusively to protect its own infrastructure.
GPT-Red works by simulating cyber attacks, something that OpenAI considers essential. The idea is to shield new generations of language models against threats before they reach the market. The main focus is to combat command injection, a trick that deceives the AI into bypassing its security rules. With assistants increasingly integrated into systems, a dangerous instruction can hide in a common email or on a web page.
How GPT-Red challenges AI security
The operation of GPT-Red is almost like that of a human hacker. It sends initial commands to the AI being tested, evaluates the response, and adjusts its messages. This cycle continues until it achieves a goal, such as forcing a data leak. To enhance its skills, OpenAI uses reinforcement learning. This means the system learns through trial and error, receiving rewards for correct actions and penalties for failures.
The result is a cycle of continuous evolution. As defense AIs become more robust, GPT-Red needs to invent more complex invasive strategies. Internal tests have shown that GPT-Red succeeds in 84% of proposed scenarios, while human experts can only overcome blocks in 13% of attempts. This already far exceeds manual work.
Practical tests and impressive results
To validate the efficiency of GPT-Red, OpenAI pitted it against real applications. In one test, it attacked the software of an AI-managed vending machine. GPT-Red managed to reduce the price of an expensive product to the minimum allowed, buying an item worth $100 for just $0.50. The loopholes were fixed after the experiment.
Another test involved a programming agent based on GPT-5.4 mini. GPT-Red easily forced the system to leak sensitive information. Additionally, it discovered a new attack tactic with a success rate exceeding 95% against the old GPT-5.1 model. Now, with the data from GPT-Red incorporated, the GPT-5.6 Sol fails in only 0.05% of direct injection attempts.
The frontier of security in artificial intelligence
Given the potential for manipulation, OpenAI decided that GPT-Red will not be released to clients or independent developers. The idea is to prevent malicious agents from using the program to automate cyber attacks. The new AI will operate behind the scenes, protecting OpenAI's infrastructure.
The creation of GPT-Red shows that OpenAI is willing to explore new frontiers to ensure the security of its technologies. This makes us wonder: how far can artificial intelligence go to protect itself?
Comments (0)
Comments are moderated and if they violate our Terms and Conditions of use, the comment will be deleted. Persistence in violation will result in a ban of your account.