Cybersecurity in a Hybrid Health-Fintech - A case of M-TIBA (White Paper Series - Part 1)

Addressing the Risk of Shadow AI in the Banking IT Environment

The rapid adoption of artificial intelligence across banking operations has created an unprecedented challenge for IT and risk leadership: the emergence of Shadow AI.  While banks invest billions in sanctioned AI initiatives, employees are simultaneously deploying unsanctioned AI tools, models, and APIs across the enterprise often without IT oversight, governance, or security controls. This parallel adoption of AI outside formal governance frameworks presents one of the most pressing operational and compliance risks facing financial institutions today.

Shadow AI is not a theoretical concern. It manifests daily in your organization: data analysts using ChatGPT for financial forecasting, customer service teams deploying third-party chatbots without security reviews, traders employing generative AI for market analysis outside approved systems, and risk teams leveraging unvetted machine learning models for credit decisions. These tools deliver immediate productivity gains, but at significant cost to data security, regulatory compliance, and operational risk management.

What Is Shadow AI in Banking?

Shadow AI refers to the unauthorized deployment and use of AI tools, models, and APIs by employees or departments without IT approval, governance review, or security validation. Unlike rogue applications of previous decades, Shadow AI operates through subscription-based SaaS platforms, public APIs, and cloud services that employees can activate with a corporate email address.

Real-World Examples in Banking Environments

• Customer Service: Departments deploying third-party chatbots to handle client inquiries, often feeding live customer account data into unvetted AI systems
• Data Analysis: Teams using generative AI for regression analysis, market forecasting, and financial modeling without model validation or audit trails
• Compliance & Risk: Risk officers using large language models (LLMs) to generate regulatory interpretation memos or draft compliance documentation
• Trading & Investment: Portfolio managers employing AI-powered sentiment analysis on unstructured market data via external platforms
• HR & Recruitment: Internal teams using AI resume screening tools to shortlist candidates without vendor security assessments
• Process Automation: Operations teams scripting workflow automation through public APIs without authentication or data governance controls

In each case, the business benefit is apparent. The governance risk is invisible until an incident occurs. 

Why Banking Is Uniquely Exposed?

Banks face a perfect storm of factors that amplify Shadow AI risk:

1. High-Sensitivity Data Environment.  Banking systems contain the most valuable and regulated data in any industry: personally identifiable information (PII), payment card data, transaction histories, credit profiles, and account credentials. A single Shadow AI incident exposing customer data can trigger regulatory enforcement, breach notification costs, reputational damage, and customer attrition. (text block 1, n.d.)

2. Stringent Regulatory Framework.  Central banks, financial regulators, and supervisory authorities have begun explicitly requiring banks to govern AI risk. Frameworks like the Basel Committee's AI governance guidance, GDPR Article 22 restrictions on automated decision-making, and emerging AI Act compliance obligations place accountability directly on the CIO and CISO.Shadow AI use particularly for credit decisions, customer segmentation, or transaction monitoring can violate these requirements unknowingly.

3. Legacy Infrastructure + Rapid Digital Transformation. Most banks operate hybrid environments: decades-old core banking systems alongside modern cloud infrastructure. This creates both pressure and permission for workarounds. Shadow AI thrives in this environment, where business units bypass legacy systems by adopting modern AI tools directly.

4. Governance Gap vs. Innovation Pressure.  IT governance frameworks often lag business demand. When formal AI governance takes 6–12 months to establish, employees find alternatives. The faster you can formalize and communicate approved AI tools, the fewer workarounds emerge.

Key Risks: Beyond Data Leakage

While data exposure is the most visible risk, Shadow AI introduces multiple operational and strategic threats:

Data Leakage & Privacy Violations

Employees sending customer PII, transaction details, or trading strategies to external AI systems creates immediate privacy violation risk. Many cloud-based AI services use customer data for model improvement unless explicitly contractually prohibited, a detail most employees never verify.

Regulatory Non-Compliance

Using Shadow AI for decision-making, credit approvals, transaction flagging, customer segmentation can violate regulatory requirements for model explainability, audit trails, and human oversight. Central banks now audit AI decision logs; Shadow AI systems typically leave no audit trail.

Model Risk & Hallucination Exposure

Large language models are known to "hallucinate" generating confident but false outputs. Using Shadow AI for regulatory interpretation, client advisory, or financial analysis can propagate false information through the organization. Who is liable when an LLM-generated memo leads to a compliance violation or client loss?

Third-Party & SaaS Risk

Every Shadow AI tool is an unvetted vendor relationship. There is no security assessment, no data processing agreement (DPA), no contractual liability clause, and no vendor financial stability check. If the vendor suffers a breach, gets acquired, or changes its terms, your organization has no recourse.

Cybersecurity & Attack Surface Expansion

Each Shadow AI deployment is a new API connection, authentication credential, and data exfiltration point. Attackers increasingly target these employee-initiated integrations because they typically bypass network security controls, endpoint detection, and data loss prevention (DLP) systems.

Detection & Visibility Strategies

You cannot manage what you cannot see. The first step is establishing comprehensive visibility into AI tool usage across the enterprise.

AI Usage Monitoring Tools

Implement monitoring solutions that log all external API calls, cloud application access, and subscription usage. Tools like Microsoft Defender for Cloud Apps or similar platforms automatically detect employee access to generative AI services, chatbots, and third-party AI platforms. Configure alerting for high-risk tools (those processing sensitive data or making decisions).

Network & Endpoint Monitoring

Deploy network detection tools to identify DNS queries and HTTPS traffic to known AI services (OpenAI, Anthropic, Cohere, etc.). Many banks are surprised to discover that network monitoring reveals dozens of Shadow AI services active at any given time. Endpoint detection and response (EDR) tools can flag API key storage, credential exfiltration, and suspicious data transfer patterns.

API Governance & Logging

Establish API gateway architecture that routes all external integrations through a centralized control point. Even if employees bypass the gateway, API logging platforms can detect the circumvention. Log all data flows to external AI services; this becomes crucial during incident investigation.

Data Loss Prevention (DLP) for AI Interactions

Extend DLP systems to monitor clipboard activity, file uploads, and data copy-paste operations targeting external AI tools. Trigger alerts when employees attempt to upload files containing PII, credit card data, or other sensitive information to cloud-based AI services.

Governance & Control Framework

Detection is necessary but insufficient. Effective Shadow AI management requires governance structures that legitimize approved AI use while controlling unapproved deployments.

AI Acceptable Use Policy

Establish a clear, specific policy that defines which AI tools are approved, which are prohibited, and which require review. Be explicit: "Employees are prohibited from uploading customer PII, account data, or transaction information to any third-party AI service without explicit CISO approval." Make the policy accessible and easy to understand; regulatory complexity is no excuse for vague governance.

Approved AI Tool Catalog

Maintain a living inventory of vetted, approved AI tools that employees can use. For each approved tool, document:

• Approved use cases (e.g., "generative AI for internal documentation drafting, not customer-facing content").
• Data classification levels permitted (public/internal/confidential).
• Authentication & access controls required.
• Contractual & security requirements met.
• Ongoing monitoring and review cadence.

This approach shifts the narrative from "don't use AI" to "here's how you use it safely."

Data Classification Rules for AI Usage

Implement tiered rules: Public data can flow to certain external AI services; internal data requires additional controls; confidential and regulated data has different rules entirely. Encode these rules into DLP and API gateway controls so they are technically enforced, not just policy-based.

Model Risk Management Integration

Shadow AI tools used for decision-making (credit, trading, compliance) must be integrated into your Model Risk Management (MRM) framework. This includes:

• Model validation and backtesting for predictive AI.
• Documentation of model assumptions and limitations.
• Regular performance monitoring and retraining assessment.
• Override procedures and human review escalation paths.

Third-Party Risk Controls for AI Vendors

Treat every approved external AI service as a vendor relationship. Require:

• Security questionnaires and data processing agreements.
• Regular security audits and vulnerability assessments.
• Financial stability monitoring (is the vendor likely to fail or be acquired?).
• Contractual liability and indemnification clauses.
• Clear data deletion and portability requirements.

Mitigation & Response Strategy

Governance creates guardrails, but operational risk requires active mitigation.

Zero Trust Approach to AI Usage

Assume that some Shadow AI deployment will occur regardless of policy. Implement Zero Trust principles for data access: require multi-factor authentication for any connection to external AI services, log all interactions, and verify data sensitivity before allowing upload. Limit employee access to approved AI tools through identity and access management (IAM) controls.

AI Gateway or AI Proxy Layer

Consider implementing a technical layer that sits between employee devices and external AI services. An AI proxy can:

• Inspect requests and responses for sensitive data.
• Block uploads of classified information automatically.
• Add audit logging to all AI interactions.
• Enforce policy-based rate limiting.
• Provide granular role-based access controls.

This is increasingly viewed as standard infrastructure in forward-looking banks.

Security Awareness & Employee Training

Shadow AI often emerges not from malice but from ignorance. Employees don't understand the risk of uploading customer data to an external chatbot. Conduct targeted training for high-risk teams:

• Customer service and support (who interface with customer data).
• Risk, compliance, and legal (whose decisions are auditable and regulated).
• Data science and analytics (who manipulate sensitive information).
• Trading and investment teams (who work with material information).

Make training specific and use real examples from your organization.

Incident Response for AI-Related Data Exposure

Develop a playbook for AI-related security incidents:

• How to detect that data has been uploaded to an external AI service.
• Chain of custody for forensic investigation (logs from DLP systems, API gateways, network monitoring).
• Vendor notification procedures and data deletion requests.
• Internal stakeholder escalation (privacy, legal, risk, customer communications).
• Regulatory notification if required.

Shadow AI incidents often escalate quickly; a pre-built response process reduces damage.

Strategic Recommendations for Banks

1. Build "AI Governance by Design"

Do not treat Shadow AI as a compliance problem to be suppressed. Instead, design governance frameworks that enable safe, legitimate AI adoption while controlling unauthorized use. 

This means:

• Rapid approval processes for low-risk AI use cases.
• Clear, transparent criteria for tool evaluation.
• Regular communication of approved tools and use cases.
• Incentives for employees to work within the governance framework.

Banks that will succeed in managing Shadow AI are those that will  make the approved path easier than the rogue path.

2. Establish a Centralized AI Oversight Committee

Create a cross-functional committee with representation from IT, Risk, Compliance, Legal, Business Units, and CISO office. 

This committee should:

• Review requests for new AI tool approvals.
• Monitor AI spending and usage trends.
• Oversee third-party vendor risk assessments.
• Coordinate incident response for AI-related breaches.
• Advise executive leadership on AI strategy and risk appetite.

Regular (monthly) meetings ensure consistent governance and rapid decision-making.

3. Balance Innovation with Compliance

The goal is not to eliminate Shadow AI through prohibition but to transform it into managed AI adoption. Aggressive prohibition drives use underground; transparent governance with clear approval paths brings it into the light. Encourage business units to formally request AI tools; make the review process fast (2–4 weeks for low-risk tools); and communicate approvals widely so others know which tools are available.

4. Invest in AI Security Tooling and Monitoring

Governance without visibility is theater. Budget for:

• SaaS and cloud application monitoring solutions.
• Enhanced DLP systems with AI-specific rule sets.
• API gateway and proxy infrastructure.
• Advanced endpoint detection tools.
• Network monitoring and DNS filtering for external AI services.

These tools represent a fraction of the cost of a regulatory fine or data breach.

Conclusion: Safe AI Adoption Without Sacrificing Innovation

Shadow AI is not a sign that your governance is failing; it is a sign that your organization recognizes the value of AI and wants to move faster than formal processes allow. The challenge for banking IT and risk leadership is to harness this energy, not suppress it.

Banks that successfully manage Shadow AI will be those that establish transparent governance frameworks, invest in visibility and control infrastructure, and create a culture where employees view approved AI tools as the path of least resistance. This requires alignment between business demand for AI capabilities and IT's ability to deliver them safely.

The regulatory environment will continue to tighten. Central bank will demand evidence of AI governance, audit trails for automated decisions, and vendor risk assessments. Organizations that wait until regulation mandates these controls will face expensive retrofitting. Those that build governance proactively will gain competitive advantage: faster innovation cycles, higher employee productivity, and lower regulatory risk.