Is it time to “protect” AI with a firewall? Arthur AI thinks so

The risks of hallucinations, personal data breaches, and regulatory compliance facing AI have raised the voices of experts and vendors that there is a clear need for some kind of protection.

One such organization currently building technology to protect against risks in AI data is New York City-based Arthur AI. Founded in 2018, the company has raised more than his $60 million to date, primarily funding machine learning monitoring and observability technology. Among the companies Arthur AI claims to be customers are three of the top five US banks: Humana, John Deere, and the US Department of Defense (DoD).

The name Arthur AI is a tribute to Arthur Samuel, widely credited with coining the term “machine learning” in 1959 and contributing to the development of some of the earliest documented models. I’m here.

Arthur AI takes AI observability one step further with today’s launch of Arthur Shield, essentially a firewall for AI data. Arthur Shield allows organizations to place a firewall in front of their Large Language Models (LLMs) to check incoming and outgoing data for potential risks and policy violations.

Arthur AI co-founder and CEO Adam Wenchel told VentureBeat: “Basically he has customers who are stuck deploying LLM, but they are stuck now, they are using this, and they are going to use this product to break the dead end.”

Does your organization need AI Guardrails or AI Firewalls?

The challenge of providing some form of protection against potentially dangerous output from generative AI is a challenge multiple vendors are trying to solve.

>>Follow VentureBeat’s ongoing generative AI coverage<

Nvidia recently announced NeMo Guardrails technology, which provides a policy language to help prevent LLMs from exposing sensitive data or hallucinating on incorrect responses. Wenchel comments that from his point of view, guardrails are interesting, but they tend to be more developer-focused.

By contrast, Arthur AI aims to differentiate itself from Arthur Shield by specifically providing tools designed to help organizations prevent real-world attacks, he said. . The technology also benefits from the observability that Arthur gains from his ML monitoring platform, which helps provide continuous feedback his loop to improve firewall effectiveness.

How Arthur Shield Minimizes LLM Risk

In the networking world, a firewall is a proven technology that filters data packets entering or leaving a network.

This is the same basic approach taken by Arthur Shield, but the prompt goes into the LLM and the data comes out. Wenchel pointed out that some prompts currently used in LLM can be quite complex. Prompts can contain user and database inputs as well as sideloading embeds.

“So we’re taking all the different pieces of data, concatenating them, putting them into the LLM prompt, and getting the response,” says Wenchel. “Additionally, there are many areas where a model can be hoaxed or hallucinated, and maliciously crafted prompts can be tricked into returning highly sensitive data.”

Arthur Shield offers a set of pre-built filters that you can continuously learn and even customize. These filters are designed to block known risks, such as potentially sensitive or toxic data, from entering or leaving LLM.

“We have an excellent research department and have really done some pioneering work in terms of applying LLM to evaluate its output,” says Wenchel. “If you upgrade your mission-critical system, you also need to upgrade the monitoring that accompanies it.”