A vulnerability in GPT-4 opened up free access to the neural network

The GPT-4 neural network is the subject of increased interest from enthusiasts and competitors
(photo: CC0 Public Domain)

A software developer named xtekky introduced a new project GPT4Free, which is a set of tools for free and almost unlimited access to a chatbot based on GPT-4 and its predecessor GPT-3.5. The exploit was created using “reverse engineering”, which helps to find a vulnerability in the neural network API.

The GPT-4 base rate is $0.03 for 1000 “hint” tokens (about 750 words) and $0.06 for 1000 “completion” tokens (also about 750 words). Access to GPT-3.5 is cheaper at $0.002 for 1000 tokens.

The GPT4Free exploit, available on GitHub, does not bypass the paid access mechanisms of the OpenAI platform, but “spoofs” the API, forcing the platform to believe that requests come from resources to which paid accounts are connected, such as the search engine You.com, the writing platform of WriteSonic text and the Quora Poe chatbot.

In its defense, the developer says, quoted by TechCrunch, that GPT4Free is intended to be run for “educational purposes only.” The enthusiast claims that he will continue to work on the project even if they try to sue him.

It is assumed that the resources that fell victim to GPT4Free will fix the security flaws and close access to the tool in the foreseeable future – GitHub can also remove the source code of the project, but this will only encourage its followers.

Access to the GPT-4 neural network is still limited, making it difficult for anyone to test it. Experts call GPT-4 ​​one of the least transparent developments of OpenAI, and the company has not yet announced when the platform will appear in the public domain.

It is likely that OpenAI will have to actively fight the “piracy” attempts of enthusiasts and competitors to get their hands on the resources of the advanced neural network.