Enterprise AI has disrupted how we do business.
But the level of efficiency we have gained has not come without trade-offs. In the race to gain a competitive edge, organizations have introduced a new security concern. Sensitive data, including account numbers, financial records, and medical information, is increasingly flowing through AI systems, while security and privacy have not kept pace. This has become one of enterprise AI’s biggest challenges today: how to use valuable data without exposing it.
Jeremy Samuelson, Chief Technology Officer at Integrated Quantum Technologies (IQT), says companies don’t have to choose between performance and privacy.
Samuelson leads the company’s innovation team at Integrated Quantum. They’ve spent a significant amount of time working to develop VEIL, a technology that transforms sensitive data into irreversible, secure representations before it enters AI model pipelines. AI models process these representations rather than the original data, reducing unnecessary data exposure while maintaining model performance (speed, accuracy, and scalability).
The research behind VEIL is backed by Dr. Mohammad Tayebi, Assistant Professor in the School of Computing Science at Simon Fraser University, a top-tier academic institution. The findings indicate the technology comes with no trade-offs to performance; it maintains predictive utility and, in some cases, improves it. This is unique compared to other security approaches that exist.
“Our focus is simple: protect the data that powers ML models without limiting their value and without willingly exposing it by using aging frameworks,” says Samuelson.
How Strong Is VEIL as a Privacy-Preserving Solution?
To evaluate the true strength of VEIL as a new privacy-preserving approach, IQT hosted a global security challenge on Kaggle, giving researchers, AI practitioners, and the data science community the opportunity to “Pierce the VEIL.” In other words, this challenge was designed to see if the irreversible representations could be reconstructed back into raw data. If the results proved that reconstruction was possible, that would mean attackers could easily steal valuable information and that the data was not well protected. If they were not able to reconstruct the representations into raw data, it would mean that even if attackers were able to hack an AI model, when VEIL is used, they would simply gain nothing of value. No raw data. No sensitive information.
“The time has come where we need to stop giving attackers any chance of getting what they want. We need to remove it instead. VEIL does that — it removes PII, PHIC, or PCI and transforms data into secure, ultra-compact representations, delivering super-charged ML models and next-gen security,” adds Samuelson.
The results of the Kaggle competition concluded that reconstructing original data from the VEIL-encoded representations was not possible.
“The fact that the Grand Prize remained unclaimed is an important validation milestone for our technology,” says Samuelson. “Just as importantly, the work produced by participants expanded our own validation efforts and reinforced confidence in VEIL’s non-invertible architecture. We are proud of the transparency of this process and grateful to everyone who participated.”
The Kaggle challenge underscores IQT’s approach to development: build, test, and validate. That discipline comes at a time when organizations face an increasingly complex threat landscape. Organizations worldwide experience an average of nearly 2,000 cyberattacks per week. Data scraping, high-profile breaches and model distillation attacks continue to rise, while growing public concern over privacy is placing additional pressure on businesses to rethink how sensitive information is handled.
That next generation of AI is introducing new questions around governance, oversight and security. To address this, IQT has initiated the patent process for MASQ (Machine Action Security Quotient), a governance and security architecture for AI agents and autonomous AI systems. MASQ is being designed to define permissions, guardrails, and boundaries for AI agents.
“AI agents are becoming increasingly autonomous and interconnected, and organizations will require governance systems capable of controlling not only what agents can access and execute, but also how sensitive contextual reasoning data is protected during machine-to-machine interaction,” adds Samuelson. “This patent initiative reflects our continued focus on building foundational infrastructure for secure enterprise AI deployment.”
