Using Anthropic AI models to attack OpenAI systems

Cybersecurity experts hacked into OpenAI systems using a software solution from archrival Anthropic. As a result, it is possible that gaps in the protection of the developers of ChatGPT were discovered and that the product itself raised issues related to cybersecurity.

Image credit: Brecht Corbeel / unsplash.com

A small group of cybersecurity experts gained access to an OpenAI employee’s ChatGPT account, allowing them to review sensitive information and recommend changes to the product’s code. The researchers used Anthropic, a tool designed specifically for experts in the field – they are paid for their work as part of a vulnerability hunting program designed to identify bugs before attackers can exploit them. The fact that experts were able to quickly hack a system at one of the world’s leading artificial intelligence labs has raised concerns about the security of artificial intelligence.

Three researchers from a small company called Hacktron AI received a $6,500 award from OpenAI for their discovery. They exploited a vulnerability in the OpenAI community forum settings to gain access to internal credentials and then accessed the ChatGPT account of one of the OpenAI employees – who had access to internal code through the GitHub platform. “We thank the researchers for contacting us and sharing their findings.”OpenAI said, adding that the vulnerabilities have been fixed.

At the same time, Anthropic reports that it has significantly increased its use of AI in the development of new models – Claude assisted in the implementation of 26% of projects in the R&D area; this figure was 1% in March alone. The company decided to make the data public to help the public. “Realizing how close the world is to a stage of recursive self-improvement” ——The stage when artificial intelligence can independently learn, self-improve, or create new models. The fear is that people will lose the ability to manage them. But it’s too early to talk about that, Anthropic adds – in 90% of tasks, AI will work with humans, although it does most of the work.