Def Con convention to host thousands of white hat hackers put against generative AI
Red team exercises will seek bugs, biases and hallucinations in large language modelsBy Alfonso Maruccia
In context: Every year, the Def Con conference brings together computer security professionals, journalists, lawyers, US federal government officials, researchers and code wizards. The target: hardware, software or anything else that can be "hacked." This year's edition will have a major focus on AI algorithms.
Scheduled for August 10-13 in Las Vegas, the 2023 edition of the Def Con conference will host a major event entirely devoted to attacking large language models (LLM) employed by generative AI algorithms. The popular hacking convention will host an AI Village (AIV) initiative, with the first public generative AI red team event.
Def Con organizers are expecting thousands of security professionals, students, and white hat hackers interested in putting LLM services to test, with AI models provided by Anthropic, Google, Hugging Face, Nvidia, OpenAI, and Stability. Microsoft will support the AI Village as well, while testers and code wizards will use an evaluation platform developed by Scale AI.
Participants will use the laptops provided by Def Con, which will also manage timed (and likely unlimited) access to multiple LLMs. A capture the flag-style point system will promote testing for a wide range of issues, with red teams expected to abide by the "hacker hippocratic oath" – whatever that actually means. The individual who will get the highest number of points will win a "high end" (albeit unspecified) Nvidia GPU.
LLMs bring an unprecedented explosion of creativity, AI Village's official blog stated, helping people getting their ideas out faster, lowering barriers to entry in creative fields, and allowing for "new kinds of creative content." However, with great powers come even greater security issues, which we are only beginning to understand as LLM technologies are being automated at scale.
Hallucinations, jailbreaks, bias, and a drastic leap in capabilities are the new issues "security professionals and the public have to deal with," AI Village said. Sven Cattell, the founder of AI Village, stated that a traditional way to solve these issues reckons on specialized red teams, mostly working in private organizations.
Def Con will provide a way to find more issues within LLM services, Cattell said, as bug bounties, live hacking events, and other "community engagements in security" can be adapted for testing against machine learning model-based systems. The August AI Village event will be a "collaborative effort," bringing together hackers, partners from community groups, non-profit groups and even (US) government supporters.