in the middle
On Thursday, the White House announced an amazing collaboration between top AI developers including OpenAI, Google, Antropic, Hugging Face, Microsoft, Nvidia and Stability AI to participate in a public evaluation of generative AI systems at DEF CON 31. bottom. Hacker convention in Las Vegas in August. The event is hosted by AI Village, a community of AI hackers.
Since last year, large-scale language models (LLMs) such as ChatGPT have become a popular way to speed up writing and communication tasks, but stakeholders recognize that they too have inherent risks. I’m here. Issues such as confabulation, jailbreaking, and prejudice pose challenges for security professionals and the general public. That’s why the White House Office of Science and Technology Policy advocates pushing these new generative AI models to their limits.
“This independent exercise will provide important information to researchers and the public about the impact of these models and encourage AI companies and developers to take steps to correct problems found in those models. We will,” said the White House statement. Compliant with the Biden Administration’s AI Bill of Rights and the National Institute of Standards and Technology’s AI Risk Management Framework.
In a side-by-side announcement written by AI Village, organizers Sven Cattell, Rumman Chowdhury, and Austin Carson called the upcoming event “the biggest red team exercise ever for any AI model group.” Thousands of people participate in public AI model evaluations using an evaluation platform developed by Scale AI.
“Red teaming” is the process by which security professionals find vulnerabilities and flaws in an organization’s systems in an effort to improve its overall security and resilience.
According to AI Village founder Cattell, “various issues with these models will not be resolved until more people know how to create and evaluate red teams.” AI Village and DEF CON aims to grow a community of researchers capable of handling vulnerabilities in AI systems by conducting the largest red-team exercises against any group of AI models.
LLM has proven surprisingly difficult to lock down, in part thanks to a technique called “immediate injection.” AI researcher Simon Willison has written more about the dangers of his prompt injection, a technique that can derail language models and cause them to perform actions not intended by their creators.
During a DEF CON event, attendees will have time-limited access to multiple LLMs via organizer-provided laptops. A capture-the-flag-style points system facilitates wide-ranging testing for potential harm. Finally, whoever gets the most points wins her high-end Nvidia GPU.
AI Village wrote: “The more people who know how to get the most out of these models and their limitations, the better the results.”
DEF CON 31 will take place August 10-13, 2023 at Caesar’s Forum in Las Vegas.
