Hackers vs. AI: The Epic Showdown Hosted by the White House and Tech Elites

How hackers at the Def Con conference tried to break AI chatbots.

Photo: AI Village @ DEF CON
In an unprecedented move, the White House recently invited hackers and security researchers to challenge some of the leading generative AI models in the industry. The challenge, part of the annual DEF CON convention in Las Vegas, saw participants attempting to deceive chatbots from tech giants like OpenAI, Google, Microsoft, Meta, and Nvidia.

Over three days, from August 11 to 13, approximately 2,200 participants took on the challenge to manipulate these large language models (LLMs) into generating misleading information, such as fake news, defamatory statements, and potentially harmful instructions. This marked the first-ever public assessment of multiple LLMs, as confirmed by a representative from the White House Office of Science and Technology Policy.

Eight tech companies, including Anthropic, Cohere, Hugging Face, and Stability AI, collaborated with the White House and event co-organizers for this initiative. To ensure fairness, the AI models were anonymized, preventing any bias towards a particular chatbot.

Kelly Crummey, a representative for the Generative Red Teaming challenge, highlighted the enthusiasm of participants, with some lining up for hours and returning multiple times. The winner, in fact, participated 21 times.

Participants at Defcon AI Hacking
Photo: Paul Bloch
Among the participants were 220 students from 19 states. Ray Glower, a computer science major from Kirkwood Community College, shared his experience with CNBC. Glower attempted various tasks, including trying to extract credit card numbers and creating a defamatory Wikipedia article. He found success in a surveillance-related task, where he managed to deceive an AI model into providing a detailed surveillance procedure.

The White House emphasized the importance of such "red teaming" exercises in identifying potential risks in AI systems. This aligns with the administration's commitment to ensuring the safety, security, and trustworthiness of AI technologies.

While the challenge's detailed results are yet to be disclosed, a policy paper is expected in October. Rumman Chowdhury, co-organizer of the event and co-founder of the AI accountability nonprofit Humane Intelligence, revealed that a comprehensive transparency report, in collaboration with the eight tech companies, will be released in February.

Chowdhury also shed light on the collaborative spirit of the event, noting that it provided a neutral space for tech companies, which often operate in silos, to come together. The challenge addressed various aspects of AI, including its internal consistency, information integrity, societal impacts, security, and more.

In a time often marked by pessimism, this collaborative effort between the government, tech companies, and nonprofits offers a glimmer of hope, signaling a collective commitment to ensuring the responsible development and deployment of AI technologies.

Read Also
Post a Comment