Hackers test chatbot security at DEF CON
LAS VEGAS, UNITED STATES — Over 2,200 participants at the DEF CON hacker convention in Las Vegas tested the security of leading chatbots, including OpenAI’s ChatGPT, during the Generative Red Team Challenge.
The goal was to identify vulnerabilities and make the bots provide incorrect information.
Kenneth Yeung, a student from the University of Ottawa, likened the process to “throwing things at a wall to see what sticks.” He emphasized that such tests help identify issues that, when addressed, can significantly improve AI systems.
The event garnered attention from the White House and major tech firms due to the potential societal implications of AI chatbot vulnerabilities.
Harvard technologist Bruce Schneier commented on the rampant vulnerabilities: “We’re just breaking stuff left and right.”
Anthropic Head of Geopolitics and Security Michael Sellitto noted that fully understanding AI safety is still a work in progress.
Recent breaches have heightened these concerns. For instance, Tom Bonner of HiddenLayer tricked a Google system into labeling malware as safe, and another expert got ChatGPT to generate malicious emails.
Furthermore, Carnegie Mellon’s research indicated major chatbots’ susceptibility to automated attacks. A separate study from a Swiss University highlighted that minimal interference can significantly corrupt AI models.
The DEF CON findings will be released in February. With AI’s growing influence, experts urge vigilance against potential security breaches and misinformation.
