Hacking contest aims to fix holes in AI models

What happens in Vegas stays in Vegas… unless it's at this weekend’s AI hacking competition in the City of Sin.

Driving the news: A hacking competition supported by big tech companies including Meta, OpenAI, and Google—as well as the US government—will play host to thousands of hackers tasked with finding vulnerabilities in the biggest AI-powered chatbots on the market. 

  • The goal of the contest is to help the tech companies at the forefront of AI fix flaws in their machine-learning models.

Catch up: The release of ChatGPT pushed Big Tech companies that were previously taking their time with AI chatbot development into a race to get their products to market, potentially skipping some testing steps in the process. 

  • After previously resisting the release of an AI-powered chatbot to the public over safety concerns, Google raced to rush to release its chatbot Bard, even as employees warned that it wasn’t ready.

  • Since then, it has created a “green lane” for its AI products, shortening the safety testing process. 

Why it’s happening: Researchers have discovered numerous vulnerabilities and problems in every major chatbot on the market, and companies are under pressure to fix them before they lead to harmful real-world consequences.

  • In spite of guardrails put in place to prevent the chatbots from providing users with dangerous information, researchers were able to trick ChatGPT, Llama-2, and Bard into providing step-by-step instructions on how to do unpleasant things like create and deploy a dangerous virus.

  • A New Zealand supermarket experimenting with AI to suggest recipes to customers has recommended meals that include deadly chlorine gas and poisonous sandwiches.

  • A recent IBM report also detailed ways to get around of large language models’ guardrails that allowed testers to get chatbots to write malicious code. 

Why it matters: AI-powered chatbots are being released and used by millions of people without their creators having a real understanding of their potential to do harm. Their ‘figure it out on the fly’ approach will be tested as AI tools become more powerful.

Bottom line: A weekend of hackers poking and prodding at these chatbots isn’t going to solve everyone's problems. As the AI race continues, these companies will have to balance safety concerns with the pressures to keep up.—LA