This AI Worm Can Steal Data And Violate ChatGPT And Geminis Security

According to a report in Wired, researchers are now creating AI worms that can steal your private information and circumvent the security measures of generative AI systems like OpenAI's ChatGPT and Google's Gemini. The first generative AI worm, known as "Morris II," was developed by researchers at Cornell University, Technion-Israel Institute of Technology, and Intuit. It has the ability to spread malware and steal data from one system to another. It was given this name in honor of the very first worm, which was released online in 1988. "It basically means that now you have the ability to conduct or to perform a new kind of cyberattack that hasn't been seen before," explained researcher Ben Nassi of Cornell Tech.

According to the outlet, the AI worm can get past some security measures in ChatGPT and Gemini by targeting a generative AI email assistant with the intention of sending spam and stealing email addresses.
To create the generative AI worm, the researchers employed a "adversarial self-replicating prompt". They claim that in response to this prompt, the generative AI model generates an alternative prompt. In order to put it into practice, the researchers integrated ChatGPT, Gemini, and open-source LLM into an email system that could send and receive messages using generative AI. They also found two ways to use the system: one was to embed the question within an image file, and the other was to use a text-based, self-replicating prompt.
In one instance, the researchers pretended to be attackers and sent an adversarial text prompt along with an email. By using retrieval-augmented generation, this "poisons" the email assistant's database and gives LLMs access to additional data from sources outside of their own. Retrieval-augmented generation, according to Mr. Nassi, "jailbreaks the GenAI service" when it retrieves an email in response to a user query and forwards it to Gemini Pro or GPT-4 for generation of a response. Eventually, this leads to data theft from the emails."The generated response containing the sensitive user data later infects new hosts when it is used to reply to an email sent to a new client and then stored in the database of the new client," he stated.