The White House is curious about what occurs when thousands of hackers come together to expose security flaws in artificial intelligence (AI) models. At DEF CON, the largest annual hacker convention in Las Vegas, major technology companies are allowing their AI systems to be tested simultaneously for the first time. The focus of the event is on large language models (LLMs) such as OpenAI’s ChatGPT and Google’s Bard. The event aims to identify problems within AI systems and foster independent evaluation.
Companies like Google, OpenAI, Microsoft, and others have agreed to make their LLMs available for hacking. The ultimate goal is to see how these models perform when challenged by determined hackers within a time limit. The event will span two and a half days, and each participant will be given 50 minutes to find flaws in eight AI LLMs. Contestants won’t be informed about the company associated with each model, although experienced hackers should be able to make educated guesses. Points will be awarded for successfully completing challenges, and the highest overall scorer will receive a powerful graphics computer as a prize. Although, the “bragging rights” associated with winning will be more valuable than the hardware.
One of the AI hacking challenges involves getting a model to invent a fact about a prominent political figure, which is intended to give AI companies more data about the frequency, conditions, and strength of LLM hallucinations. The event is supported by the White House, which hopes to obtain critical national security information about these AI models and to foster improvements, especially in light of looming AI disinformation campaigns.
The event, which looks to become an annual DEF CON AI Village activity, is focused on exposing glitches, bugs, and correctable AI issues, rather than focus on existential threats. The goal is to determine if AI models contain biases, disseminate false information, or invent fictitious details. Event organizers have urged government intervention to regulate AI and combat misinformation. Once the event concludes, the participating AI companies will review the collected data, and address or patch any identified flaws. Independent researchers will also have the opportunity to access the contest data; the results will be published in February.
The ainewsarticles.com article you just read is a brief synopsis; the original article can be found here: Read the Full Article…
