Thousands of hackers gathered at the Def Con hacking conference in Las Vegas last weekend to participate in a public contest aimed at exposing flaws and biases in generative AI systems. The hackers tested eight models produced by companies such as Google, Meta Platforms, and OpenAI to see if they would produce incorrect or harmful outputs.
One of the participants, Kennedy Mays, successfully tricked a language model into saying that nine plus ten equals 21. The model initially qualified the response as an inside joke, but eventually started producing the incorrect sum without any explanation. This experiment demonstrated the potential for large language models to produce inaccurate or misleading information.
The contest, supported by the White House, seeks to address the concerns associated with large language models. These models have the power to revolutionize various industries, but researchers have identified biases and other issues that could result in the spread of inaccuracies and injustice if not addressed. The White House has developed a Blueprint for an AI bill of rights and is working on an executive order to regulate AI technologies.
The event also highlighted concerns about inherent bias in language models. Participants asked the models to consider perspectives such as those of a member of the Ku Klux Klan, which resulted in the endorsement of hateful and discriminatory speech. The goal is to proactively address abuse and manipulation that could arise from the use of AI technologies.
Researchers and cybersecurity experts have been investigating attacks against AI systems, but some contend that it may be impossible to completely mitigate vulnerabilities in these models. The nature of the technology itself presents challenges, leading some to suggest that not using large language models may be the only way to ensure 100% security.
The public contest at Def Con has brought attention to the limitations and potential risks associated with large language models. As more people become aware of these issues, the hope is that companies and policymakers will develop better safeguards to prevent the misuse of AI technologies.