A recent Fortified webinar poll shows how quickly organizations are using AI in healthcare. Eighty-two percent of participants said their organizations were piloting AI, using it across multiple workflows, or relying on tools with AI already built in.
As AI becomes part of more everyday tools, security teams have more to find and track. A feature added by a trusted vendor may reach new data, connect to another system, or take action that wasn’t part of the original security review.
That is where cybersecurity and patient care risks can change silently without knowledge or consent.
Healthcare Impact
Newly incorporated AI components can change a vendor’s risk profile long after the initial assessment. The vendor may still be approved, and the product may still be considered business essential. But what the product can now do with AI enablement may have changed behind the scenes. Security and third-party risk teams need a way to catch those changes, assess that risk to data exposure, sub-processing, and data use to include retention.
Healthcare teams across the organization need to review what the new AI capability can access, which systems it connects to, what permissions it has, and what actions it can take. Review whether those changes affect patient data, clinical workflows, identity, privacy, data exposure or incident response capabilities.
Leadership Recommendations
Reassess AI after adoption
Do not let the original approval become the last review. Build reassessment into procurement, change management, privacy, security, clinical review, and third-party risk. When a vendor adds a major AI feature or changes how an existing one works, plan for the reassessment and review it again.
Match oversight to consequence
Not every AI use case needs the same level of review. Look at the data it can access, the decisions it can influence, the actions it can take, and the impact if something goes wrong. Apply the strongest controls where AI can affect patient care, patient communication, security, operations, or finances.
Test the human in the loop
A human review step only works if the person reviewing the output can challenge it. Define who reviews the output, what information they need, when independent verification is required, and when the process should stop or escalate. Does the human in the loop have the ability to provide feedback to modify future outputs and how are those changes tracked?
Watch for influence, not only automation
AI does not have to make the final decision to shape the outcome. Recommendations, summaries, rankings, and suggested next steps can influence what a person sees first and which options they consider. Review how those features affect judgment, especially when the consequences are significant.
Essential Healthcare Leadership Cybersecurity Questions:
- Has the AI we approved changed since the original assessment?
- Where is AI assisting people, where is it influencing important decisions, and where can it act?
- Does the level of oversight match what the capability can access and do?
- When we say a person reviews an AI output, what does that review require?
- Who can stop or escalate an AI-enabled process when its behavior, access, or scope changes?
Executive Takeaways:
- AI risk can change after implementation
- Trusted vendors still need to be reassessed when AI capabilities change
- Human involvement only works when the review is meaningful
- Oversight should reflect what the AI can access, influence, and do
THREAT NEWS AND INSIGHTS
- Rotate Now: OpenCode Console Breach May Have Exposed OpenAI, Anthropic and Google API Keys
- Critical PHP2Shell Flaws Expose PHP Web Servers to Unauthenticated Remote Takeover
- Emergency Triage and Patch Required: Two Citrix NetScaler Zero-Days Under Global Attack
Peer Pulse: Finding the AI You Did Not Know You Had
In this month’s Peer Pulse, Russell Teague speaks with Dr. Murali Naidu, CMO, Vitea, about shadow AI, why it is difficult to discover, and the cybersecurity risks created when AI operates outside established security and governance processes.
Russell Teague: When you work with healthcare organizations, how much AI use do you believe security teams know about?
Dr. Murali Naidu: Most organizations know about the AI they intentionally deployed. The bigger visibility challenge is AI entering through individual users, vendor integrations, embedded features, APIs, and tools not introduced as stand-alone AI products. In our experience at Vitea, clients are surprised to find that they have 6-8x as many AI tools operating on their network.
Russell Teague: Where are healthcare organizations most likely to discover AI they did not intentionally deploy or approve?
Dr. Murali Naidu: Look beyond the official AI inventory. Shadow AI can appear through employees using personal tools, departments adopting services independently, vendors adding functionality, and AI becoming embedded in browsers, applications, and clinical workflows.
Russell Teague: When you discover shadow AI, what cybersecurity risks concern you first?
Dr. Murali Naidu: Start with the data. Determine what information entered the tool, where it went, whether it was retained, and who can access it. Then examine the identities, permissions, integrations, and actions connected to that capability. My strong suggestion is to start with tools that interact with patients or patient information directly, as that is often the highest-risk category.
Russell Teague: Is the larger risk an AI tool selected by an employee or AI functionality added to technology the organization already trusts?
Dr. Murali Naidu: Both can create exposure, but embedded AI may be easier to overlook because the underlying product is already trusted. A familiar vendor or application can gain new access, integrations, or automation without receiving the same scrutiny as a new technology. That said, we have created an environment of access (for good reason), moving away from organization-provided devices. Therefore, we have ubiquitous use of personal devices through which people can access health system information and tools while on a cellular network.
Russell Teague: What makes AI discovery different from a traditional software or asset inventory?
Dr. Murali Naidu: A point-in-time product inventory tells you what the organization owns. AI discovery must also identify capabilities operating inside those products, including agents, APIs, vendor integrations, automated workflows, and the data and permissions connected to them. These tools are dynamic. Not just through updates. They are literally changing constantly.
Russell Teague: If a CISO believes the official inventory understates AI use, where should the organization look first?
Dr. Murali Naidu: Start with vendor applications, browser-based services, productivity tools, clinical integrations, API activity, and workflows handling sensitive information. The goal is not only to identify the tool, but to understand what it can access, influence, and do.
CLOSING PERSPECTIVE
Discovery and inventory of AI within healthcare organizations will be critical moving forward.
Healthcare leaders need to know when new AI capabilities alter what a trusted tool can access, connect to, or do. Find the change, understand the exposure, and reassess the risk before it affects cybersecurity posture or patient care.
