In this article, the author discusses how large language models (LLMs) are being used to process and generate human-like text. These models, such as ChatGPT and Google Bard, have the ability to answer questions, translate languages, and summarize text. However, there is growing concern about their ability to generate objectionable content. Researchers at Carnegie Mellon University have discovered a vulnerability that allows them to manipulate LLMs into producing objectionable behaviors at a high rate of success. This is concerning because as autonomous systems become more prevalent, it is important to prevent them from being hijacked by these kinds of attacks. The researchers are now focused on finding ways to fix these vulnerabilities.
https://www.cmu.edu/news/stories/archives/2023/july/researchers-discover-new-vulnerability-in-large-language-models