Keeping Data Private and Secure with Agentic AI

When it comes to data privacy and AI, companies are in a tough spot. On the one hand, businesses are eager to take advantage of technological advances in AI, including the development of autonomous AI agents. But on the other hand, the potential risks around data leakage and violating data regulations are putting a damper on the AI enthusiasm. The folks at confidential computing startup Opaque say a new release of their platform could provide a solution.

Opaque is an open source confidential computing project that emerged nearly a decade ago at RISELab, the UC Berkeley computer science lab that succeeded AMPlab and preceded the current Skylab. In 2021, several RISELab participants co-founded Opaque (the company), including RISELab directors Ion Stoica and Raluca Ada Popa, Professor Wenting Zheng, and RISELab grad students Rishabh Poddar and Chester Leung.

As a confidential computing project, Opaque provides certain guarantees around the security and the privacy of data that is processed within its framework. The original confidential computing work centered on the Multiparty Collaboration and Competition (MC2) platform, which enabled multiple data owners to perform joint analytics and ML model training on collective data without revealing their individual data to each other.

Today, Opaque is offering a confidential computing platform where customers can build and run their AI applications with full data privacy and security guarantees. Customers that use Opaque’s platform) get built-in encryption of data, encryption key management, column- and row-level access control, and tamper-proof audit trails, among other capabilities.

GenAI Holdups

The potential impact of GenAI is huge. A 2023 study by McKinsey concluded that the tech could add $2.6 trillion to $4.4 trillion to the world’s economy every year. Despite the massive potential, only a small fraction of GenAI applications are actually making it out of the development and testing phase. Numerous surveys of companies have highlighted security and privacy as primary reason for this GenAI holdup.

Opaque uses confidential computing techniqees to keep data secure in GenAI workflows (Image courtesy Opaque)

For instance, a 2024 Dataiku study identified that the biggest concerns around GenAI are a lack of governance and usage control, cited by 77% of the survey respondents. Cloudera’s State of Enterprise AI and Modern Data Architecture report concluded that the top barriers to adopting AI were worries about the security and compliance risks that AI presents (74%). And a 2024 IBM Institute for Business Value study found that 80% of CEOs said transparency in their organization’s use of next-generation technologies, such as GenAI, is critical for fostering trust.

The guarantees provided by Opaque should help companies move their AI applications from the development and testing phase into production.

“The core value proposition of Opaque is we’re helping companies accelerate their AI into production,” says Leung, the head of platform architecture for Opaque. “It enables data to be used for machine learning and AI without compromising on the privacy and the sovereignty of that data.”

Companies with advanced encryption skills could potentially build their own confidential computing frameworks that provide the same privacy and security guarantees as Opaque, Leung says. However, folks with those skills are typically not widely available on the open market, particularly when it comes to building large-scale, distributed applications used by large enterprises, which is Opaque’s target market.

“Confidential computing requires you to understand cryptography. It requires you to understand systems and how to mess with the systems in a way that will keep them secure, and that will allow you to scale them,” Leung tells BigDATAwire in an interview. “All of that knowledge is not really that accessible to an everyday data scientist…It’s not the easiest thing to pick up, unfortunately.”

Transparency and Opacity

Following the development of MC2, the San Francisco-based company’s first commercial product was a gateway that sat between the GenAI application and the third-party large language model (LLM), and prevented sensitive data contained in the GenAI prompts and retrieval augmented generation (RAG) pipeline from leaking back into the LLM.

Its latest offering supports emerging agentic AI architectures and provide security guarantees on data and workflows that span multiple systems.

The Opaque co-founders, left to right: Leung, Poddar, Ada Popa, Zheng, and Stoica (Image courtesy Opaque)

“Traditionally, we’ve been focused on kind of batch analytics, batch machine learning jobs,” says Leung, who’s advisor at RISElab was 2023 BigDATAwire Person to Watch Raluca Ada Popa. “We later then supported kind of more general AI pipelines, and now we’re building specifically for agentic applications.”

Opaque, which has raised $31.5 million in seed and Series A money, is targeting big Fortune 500 firms that want to roll out AI-powered applications while navigating strict data regulations and complex back-office systems. For instance, it’s helping the SaaS vendor ServiceNow develop a help desk agent that can handle sensitive data without violating privacy guidelines.

In the ServiceNow case, sales reps may have questions about how their commissions are calculated. The challenge for the autonomous AI agent is that it must have access to and process a variety of sensitive data, such as annual contract values and private financial data, to explain to the sales reps how their commissions were calculated.

“We provide what we’re calling this confidential genetic architecture for them to use as the back end for their employee help desk agent,” Leung says. “They’re relying on us to power the security, privacy side of things.”

As more companies begin to develop agentic AI systems, they may find Opaque’s new Compound AI for Agents architecture handy to solve thorny security and privacy issues. According to Opaque, the new agentic AI architecture will ensure “that every aspect of agent reasoning and tool usage maintains verifiable privacy and security.”

More Data, Please

AI is fundamentally a product of data. Without high quality data to train or fine-tune an AI model, the odds of building a good model are somewhere between slim and none. And while the amount of data the world is generating continues its upward trajectory, data scientists are finding that they have less access to data, not more. Leung hopes that confidential computing will turn that trend around.

“Advancements have created this huge demand for data,” he says. “The more data you have, and in particular, the more high quality data you have, generally the better your AI is. That’s true for traditional AI. That’s true for generative AI.

“Now, what we’ve been seeing over the last decade…is that the supply of high-quality data has actually gone down, because the data is fragmented, because regulations, risk teams, and legal teams are placing restrictions on how you can actually use that data,” he Leung continues.

That’s created a tension between the supply of data and the demand–a tension that could potentially be resolved with confidential computing technologies and techniques. Opaque certainly isn’t the only company chasing that dream, but considering the decade that it’s already spent working on the problem with some of the top computer scientists in the country, it should be considered one of the early leaders in this emerging space.

RISELab Replaces AMPLab with Secure, Real-Time Focus

Yes, You Can Do AI Without Sacrificing Privacy

The post Keeping Data Private and Secure with Agentic AI appeared first on BigDATAwire.