Using a publicly available large language model, such as OpenAI’s ChatGPT, comes with certain security risks. Any data you might share with a large language model (LLM) could be stored in the chat history and may then be used for training, helping the model answer future prompts from other users. As such, LLM users who share sensitive information with a model – say unpublished data or anything confidential – may run the risk of that information being leaked later on.
“If you tell an OpenAI model, ‘Review my article and provide commentary and suggestions for improvements,’ that prompt may be used to train the model later on,” said Irene Kopaliani, a cloud architect with Princeton Research Computing. “So inadvertently, you could share more information with it than you expected to.”
Kopaliani approves and processes user requests for the Princeton AI Sandbox – a new tool from the Office of Information Technology in collaboration with research computing which aims to reduce the security risks posed by using LLMs. The sandbox provides a secure environment for Princeton University researchers to explore and use models for their projects. “Unlike public LLMs, our sandbox does not use user data for model training or share it with anyone,” said Kopaliani. “Chats are automatically deleted after 7 days and are only stored temporarily for user convenience.”
The concept of sandboxes, or secure and controlled environments in which users can explore the use of AI, are not unique to Princeton. In the last several years, institutions have begun implementing sandbox environments for their researchers, including Harvard University and the University of Michigan.
Seeing that other universities had secure AI environments, faculty at Princeton began requesting a similar tool, said Mark Ratliff, director of cloud infrastructure services at Princeton. Ratliff was a part of the team which ultimately developed and then launched Princeton’s AI Sandbox. The first user was given access in February.
“The thing really that this service provides for Princeton researchers is a stamp of confidence saying that we had reviewed the environment and it’s safe,” said Ratliff. Otherwise, the features of the tools available in the sandbox remain largely the same as they would be if accessed online instead. “We simply make use of the AI services that Microsoft manages.”
Users of the AI Sandbox are provided unlimited access to the models via a web-based chat interface and via a programming API. Requests for an account with the AI Sandbox must come from a Princeton faculty member via a request form; however, researchers including undergraduates, graduate students, or postdocs can get access by faculty sponsorship.
The security benefits of the AI Sandbox are its main selling point, but other users on campus find there are more than just one advantage to using the tool.
Benedikt Stroebl, a Ph.D. candidate being advised by Professor of Computer Science Arvind Narayanan, has been using the AI Sandbox for his research on AI agents. “The AI Sandbox is great because it comes basically at no cost for you, in terms of financial cost, but also set up cost,” said Stroebl. “That’s very useful for many Ph.D. researchers, especially junior researchers, who sometimes may not have the funding to cover those costs.”
Stroebl’s focus is on evaluating AI agents, which are often autonomous software programs that are capable of accomplishing complex tasks such as software engineering or web research. In his research, Stroebl and his collaborators use the models that power AI agents and run them on various benchmarks. This often leads to a large accumulation of inference costs – the costs which arise from using a trained model to generate outputs. “One big advantage is that I have access to these models through an API that is free to use,” said Stroebl. He’s been using the sandbox to run the evaluations across different models.
Overall, Stroebl emphasized his positive experience using the tool and recommended it as a useful resource for other researchers on campus. “The AI sandbox is a great initiative and we’re very lucky at Princeton to have these sort of resources available for our research,” said Stroebl. “I'm just happy and excited that something like the Sandbox is available.”
If you’re interested in using the AI Sandbox, you can access more information here: https://researchcomputing.princeton.edu/support/knowledge-base/ai-sandbox