Researchers from Carnegie Mellon University and the Center for Artificial Intelligence Safety in San Francisco claim to have identified a way to consistently bypass security measures in AI chatbots like ChatGPT and Google Bard, allowing them to generate harmful content. The study, reported on July 27, reveals that appending long-character suffixes to prompts entered into chatbots can circumvent the anti-abuse protection systems used to prevent the generation of hate speech, disinformation, and toxic material.
The researchers conducted tests, including requesting a bomb-making tutorial from a chatbot, which the AI refused to provide. While the companies behind large language models can block specific suffixes, they currently lack a comprehensive solution to prevent all such attacks, leaving room for potential misuse. Concerns have arisen that AI chatbots could flood the internet with dangerous content and misinformation, and addressing this issue has become crucial.
Zico Kolter, a professor at Carnegie Mellon University and an author of the report, stated that there is no clear-cut solution, as numerous attacks can be launched in a short period of time. The research was shared with AI developers Anthropic, Google, and OpenAI, eliciting responses from these companies. OpenAI spokesperson Hannah Wong expressed appreciation for the research and stated that they have been actively working to enhance their models' robustness against adversarial attacks.
Somesh Jha, an AI security specialist and professor at the University of Wisconsin-Madison, commented on the potential consequences of such vulnerabilities. If these issues persist, governments might be prompted to enact legislation aimed at regulating AI systems. The study underscores the need to address risks before deploying chatbots in sensitive areas, as vulnerabilities could have severe implications if not addressed proactively.
In May, Carnegie Mellon University received significant federal funding of $20 million to establish a new artificial intelligence institute aimed at shaping public policy, showing the growing importance of AI safety and governance.
