United States-based researchers have claimed to have found a way to consistently circumvent safety measures from artificial intelligence chatbots such as ChatGPT and Bard to generate harmful content.
According to a report released on July 27 by researchers at Carnegie Mellon University and the Center for AI Safety in San Francisco, there’s a relatively easy method to get around safety measures used to stop chatbots from generating hate speech, disinformation, and toxic material.
Well, the biggest potential infohazard is the method itself I suppose. You can find it on github. https://t.co/2UNz2BfJ3H
The circumvention method involves appending long suffixes of characters to prompts fed into the chatbots such as ChatGPT, Claude, and Google Bard.
The researchers used an example of asking the chatbot for a tutorial on how to make a bomb, which it declined to provide.
Researchers noted that even though companies behind these LLMs, such as OpenAI and Google, could block specific suffixes, here is no known way of preventing all attacks of this kind.
The research also highlighted increasing concern that AI chatbots could flood the internet with dangerous content and misinformation.
Professor at Carnegie Mellon and an author of the report, Zico Kolter, said:
The findings were presented to AI developers Anthropic, Google, and OpenAI for their responses earlier in the week.
OpenAI spokeswoman, Hannah Wong told the New York Times they appreciate the research and are “consistently working on making our models more robust against adversarial attacks.”
Professor at the University of Wisconsin-Madison specializing in AI security, Somesh Jha, commented if these types of vulnerabilities keep being discovered, “it could lead to government
Read more on cointelegraph.com