Researchers Find 99.8% Exploit in Meta's Prompt-Guard-86M AI Model

Aug 2, 2024, 03:54 PM

Researchers at robusthq have identified a significant vulnerability in Meta's recently refreshed Prompt-Guard-86M model, which is designed to protect large language models (LLMs) against jailbreaks and other adversarial examples. The exploit has a 99.8% success rate. The researchers have shared countermeasures with Meta, and the company is working on a fix. The findings were published in a blog. Additionally, a new method has been developed to enhance the security of open-source LLMs by preventing tampering, which could prevent misuse such as explaining how to make a bomb.