Artificial intelligence is transforming enterprise operations, but much of that transformation is happening beyond the visibility of corporate security teams. Third-party vendors are adding AI features, integrating external models and deploying autonomous agents, often without notifying their customers.
In a recent SC Media webcast, host Joshua Marpet and Demi Ben-Ari, co-founder and chief strategy officer at Panorays, explored how the proliferation of third-party shadow AI is undermining traditional vendor risk management.
Their central concern was the speed of change. For example, a vendor might pass its annual security assessment, then introduce AI functionality months later through a routine software update.
Marpet explained why that creates a dangerous gap.
"It's not a new product that you've got to do a reassessment on," he said. "It's a feature upgrade on the existing products that your vendor risk assessment is fine for."
If that keeps happening, the result will be a constantly expanding network of AI-enabled suppliers, models and integrations that client organizations may not even know exist, but which threaten their security postures nonetheless.
Ben-Ari and Marpet both lamented that the speed at which organizations are implementing AI is outstripping today's security controls.
"Public companies right now are measured upon AI adoption," Ben-Ari said. "Everything is happening at hyperscale … everything that you've implemented in the matter of security, privacy, risk controls, etc., might become obsolete."
Why traditional vendor assessments fall short
Conventional third-party risk management relies heavily on manual questionnaires, standardized certifications and periodic, often yearly, assessments. Those mechanisms are still useful today, but they adequately capture the reality of constantly changing AI deployments.
Ben-Ari argued that organizations must reconsider assessment frequency.
"Right now, point-in-time assessments are still the process itself, but the cadence in which it needs to happen are not point-in-time anymore," he said.
Ben-Ari recommended implementing continuous or event-triggered monitoring that can identify new AI features, changed permissions, additional integrations and emerging data-processing relationships.
Importantly, the unit of assessment should not always be the vendor itself. A single software provider might offer project management, code repositories and knowledge-management applications. AI capabilities enabled in one service may create risks that do not exist in another.
To compensate for this, organizations will need to evaluate their individual business relationships, including which applications process sensitive information and how AI might change those interactions.
Following data through the AI supply chain
Ben-Ari and Marpet repeatedly returned to the importance of understanding where enterprise data travels. An AI-powered sales application, for example, might connect to a CRM system, send customer information to an external model and invoke additional services through APIs or agents.
Those relationships become even more complicated when a vendor relies on fourth-party providers, which themselves depend on additional external AI services.
Marpet described emerging agent-to-agent workflows in which autonomous systems exchange information across multiple suppliers, potentially without the customer's knowledge.
Ben-Ari recommended mapping these relationships to identify data flows, permissions, model dependencies and potential attack paths. He also emphasized identifying the organization's most valuable data and business-critical processes before deciding which third-party relationships require the greatest scrutiny.
But Marpet lamented the sheer scale of information that must be processed to gain total visibility into AI operations.
"t's tough because there's browser-based AI. There's installed-on-my-local machine AI. There's AI I use through an API. There's agents. There's all kinds of things going on," he said. "It's extraordinarily difficult to collect enough information in the right ways. You have to have a huge application to get all that, and it has to do 40 different things."
Of course, not every vendor presents equal risk. For example, a supplier handling routine office purchases warrants having different, presumably looser controls than an AI integration with privileged database access.
Building continuous, risk-based AI governance
Both speakers discouraged relying exclusively on blanket restrictions against particular AI models or providers.
"These kinds of policies — we can only use OpenAI, we can only use Copilot, we can only use this — are effectively relics," Marpet said. "It is not the model you should be mandating. It is the behaviors."
That means examining data access, processing locations, permissions, agent capabilities and business purpose.
Ben-Ari recommended establishing a cross-functional AI committee involving security, privacy, legal, compliance, procurement and business stakeholders.
He also advocated using established frameworks, including NIST guidance and ISO/IEC 42001, as starting points for risk-based assessments rather than attempting to build an entirely new governance structure.
Automation is essential. Organizations may maintain thousands of supplier relationships, making manual monitoring impractical. Telemetry from browsers, procurement systems and third-party risk platforms can help identify previously unknown AI activity and trigger reassessments when circumstances change.
Equally important are basic operational details: knowing who owns each vendor relationship, who handles the associated data and whom to contact when an incident occurs.
From shadow AI to accountable AI
The discussion's overarching message was that third-party AI risk clearly cannot be managed through annual questionnaires and static inventories.
Organizations must be able to continuously discover AI-enabled relationships, understand how sensitive data moves through them and assess whether changing capabilities introduce unacceptable exposure.
Ben-Ari concluded with three priorities: Establish an appropriate governance framework, assemble the stakeholders responsible for AI oversight, and collect the telemetry necessary to support informed decisions.
The objective is not to stop AI adoption, he explained. It is to make third-party AI visible, measurable and accountable before an unexpected feature update, compromised integration or autonomous agent creates a security incident.
