The rapid adoption of AI across the enterprise has transformed how organizations build software, automate workflows, and improve productivity. But as AI becomes increasingly embedded in everyday business operations, it is also exposing security gaps that many organizations are unaware of. Traditional methods for understanding software components and managing supply chain risk are no longer equipped to deal with systems that can interpret information, make decisions, and act independently.
In this episode of UC Today, Kristian McCann is joined by Devashri Datta, Senior Security Architect at NVIDIA, who spoke to us in a personal capacity as an independent researcher specializing in AI supply chain security and open source governance. Drawing on her research, Datta explains why enterprises need to rethink how they approach software transparency and governance as AI becomes more deeply integrated into critical business systems.
Why Traditional SBOMs Fall Short in the AI Era
Datta begins by explaining the original purpose of an Software Bill of Materials (SBOM), describing it as the software equivalent of an ingredients label on food packaging. It identifies the components that make up an application, including third-party libraries, dependencies, licenses, and versions, giving organizations greater visibility into what their software contains.
Over the past two decades, SBOMs have become increasingly sophisticated. Machine-readable standards such as SPDX, vulnerability extensions like CycloneDX, and a new VEX (Vulnerability Exploitability eXchange) have all improved the way organizations identify and prioritize software vulnerabilities. However, according to Datta, those improvements were designed for traditional software rather than AI-powered systems.
"We are moving from an era that says we execute systems to code," she explains.
"Now we are moving from systems to code to an AI system that interprets intent and takes actions."
That distinction fundamentally changes what organizations need to monitor. Instead of tracking only software components, AI introduces entirely new elements, including model weights, training data, retrieval-augmented generation (RAG) pipelines, fine-tuning layers, and integrations with multiple external tools. Traditional SBOMs were never designed to capture this information, leaving significant blind spots in software governance.
Datta also highlights that AI introduces risks that simply do not exist in conventional applications. Prompt injection, data poisoning, hidden instructions embedded within documents, and tool misuse all create attack vectors that cannot be identified simply by scanning software packages. Security teams must now think beyond code vulnerabilities and consider whether an AI system itself can be manipulated into performing harmful actions.
Building Governance Around AI Behavior Rather Than Code
Those risks become even more significant as AI assistants gain access to enterprise collaboration and productivity platforms.
Datta points to widely used tools such as Microsoft Copilot, Cursor, Claude, Gemini, and Codex, explaining that these platforms increasingly interact with emails, Teams or Slack conversations, documentation, Jira tickets, GitHub repositories, PDFs, and internal knowledge bases. Rather than acting as passive assistants, they are beginning to execute tasks and make decisions on behalf of employees.
