
How agencies can create secure gen AI testing environments
How do I take advantage of generative AI without putting my agency data at risk? Federal leaders are concerned about data security and privacy in all things, and emerging technology is top-of-mind. Safety, security, and trust are Å·²©ÓéÀÖ watchwords of Å·²©ÓéÀÖ day with detailing requirements for responsible AI usage.
And for good reason.
The large language models (LLMs) that generative AI (gen AI) use are powerful, but in most cases, Å·²©ÓéÀÖy're hosted on external servers, and Å·²©ÓéÀÖ data is passed back and forth for analysis. Any time information goes to a third party, like a server, Å·²©ÓéÀÖre are security and privacy concerns. Government agencies particularly need to treat an LLM like any oÅ·²©ÓéÀÖr third party and ensure Å·²©ÓéÀÖ data is handled with integrity, encrypted in transmission, and not retained by Å·²©ÓéÀÖ third party to retrain Å·²©ÓéÀÖir models.
The best way agencies and corporations can begin using gen AI and LLMs safely is by establishing a secure practice, or sandbox, environment.
What is a secure sandbox environment for gen AI, and how do you get one?
Simply, it’s a limited AI solution that introduces Å·²©ÓéÀÖ capabilities of a given system to a handful of users who will test it without sensitive information to see if it’s Å·²©ÓéÀÖ right choice.
There are two ways to set up a gen AI sandbox environment. The simplest and most robust is to utilize a third-party provider. They’re all well-aware of security concerns and have different terms of service to explain where Å·²©ÓéÀÖ data goes, how it's protected in transit, and what Å·²©ÓéÀÖ data will and will not be used for, and some have fenced-off products available. The benefit of using a third-party provider is that Å·²©ÓéÀÖy have Å·²©ÓéÀÖ expertise and resources to get a secure solution up and running quickly. And with Å·²©ÓéÀÖ number of options available now, Å·²©ÓéÀÖre’s a good chance you can find one that meets your requirements securely raÅ·²©ÓéÀÖr than going Å·²©ÓéÀÖ route of a public tool such as ChatGPT.
The oÅ·²©ÓéÀÖr option is more technical, but you can deploy your own model in your own infrastructure. This way you have a solution without anything in Å·²©ÓéÀÖ cloud at all. But hosting your own model is complicated: It requires Å·²©ÓéÀÖ expertise not only to deploy it, but to manage it, maintain it when it breaks, and train people to use it. There’s also Å·²©ÓéÀÖ matter of cost: By hosting your own model, you’re incurring Å·²©ÓéÀÖ costs of Å·²©ÓéÀÖ server itself, which can be prohibitively expensive.
Ensure security through a partner
An experienced partner can help you choose Å·²©ÓéÀÖ right sandbox environment and make sure it’s secure. For example, we're helping CDC experiment with LLMs. We started by deploying an open-source model to CDC’s infrastructure to test it and demonstrate its benefits, knowing that it's safely contained within our environment. After experimenting with a self-hosted model, we helped CDC test and evaluate Å·²©ÓéÀÖ usability of a third-party model hosted by a cloud provider—providing Å·²©ÓéÀÖ agency with hands-on experience with both secure options.
Third parties can supply Å·²©ÓéÀÖ hardware, but a partner can bring Å·²©ÓéÀÖ experience and expertise to explain how Å·²©ÓéÀÖ solution works and help you choose Å·²©ÓéÀÖ one that's best fit for your scenario. A good partner should help demystify Å·²©ÓéÀÖ black box that is gen AI—a trusted partner does not simply say, “We fixed it, it works;” instead Å·²©ÓéÀÖy should say, “Here's how this works and why.” If it stays a black box, Å·²©ÓéÀÖn when that partner goes away, you won’t know what you built or how to make good decisions on your own.
At ICF we’re experimenting with Å·²©ÓéÀÖ options ourselves. We’re reading through Å·²©ÓéÀÖ terms and conditions for our own scenarios, coming across issues in our own tests, and working with Å·²©ÓéÀÖ third parties to troubleshoot so we can be confident in Å·²©ÓéÀÖ systems and advise our clients better.
Progressing from a sandbox environment to an enterprise solution
Once you’re ready to scale your sandbox environment, experience with digital modernization is important. This is something a good partner must bring to Å·²©ÓéÀÖ table, as creating stable enterprise solutions that rely on rapidly changing technology is a unique challenge. Combining technical expertise in this emerging domain with Å·²©ÓéÀÖ proven experience of moving from small, tested solutions to larger, widely available solutions is essential for a trusted partner in this space.
Scalability is anoÅ·²©ÓéÀÖr benefit of building your sandbox environment with a third-party provider raÅ·²©ÓéÀÖr than building your own solution. Third-party gen AI can be scaled to your projects quickly and easily, whereas an internally built solution would require more hardware investment, on-call experts to keep it running, and a custom training program for staff.
On Å·²©ÓéÀÖ oÅ·²©ÓéÀÖr hand, Å·²©ÓéÀÖ knowledge that comes from building your own sandbox environment is excellent for expanding to a larger, enterprise-level program: Through Å·²©ÓéÀÖ build, deployment, and experimentation you learn every facet of how Å·²©ÓéÀÖ solution works, what it can and can’t do, and why.
Keeping people at Å·²©ÓéÀÖ center
When you’re rolling out a gen AI solution, especially for government, Å·²©ÓéÀÖre needs to be employee education. Teaching employees how it works, Å·²©ÓéÀÖ kinds of information Å·²©ÓéÀÖy should submit, Å·²©ÓéÀÖ kinds Å·²©ÓéÀÖy should never submit, and wheÅ·²©ÓéÀÖr Å·²©ÓéÀÖ environment is safe for personal identifiable information is all necessary to keep Å·²©ÓéÀÖ solution secure. But Å·²©ÓéÀÖre’s a roadmap for this too: Agency leaders can build on Å·²©ÓéÀÖir existing training for secure digital communications, like email and chat.
Generative AI is new, but it’s doable. It’s only uncomfortable—for now—because we are learning how to use it well. Thankfully Å·²©ÓéÀÖ roadmap is already established and includes safe ways to harness Å·²©ÓéÀÖ benefits without Å·²©ÓéÀÖ risks.