Securing AI: Navigating risks and compliance for the future

By Yonatan Zunger, Deputy CISO and CVP of AI Safety and Security, Microsoft

AI is reshaping industries, revolutionizing workflows, and driving real-time decision-making. Organizations are embracing it at an astonishing pace. In fact, 47% of AI users already trust it to make critical security decisions.¹ That’s a clear sign that AI is becoming an essential force in business. But here’s the challenge—if not secured properly, AI’s immense potential can become a setback to deploying AI across your organization.

As AI becomes more deeply embedded in workflows, having a secure foundation from the start is essential for adapting to new innovations with confidence and ease. New regulations like the European Union AI Act demand greater transparency and accountability, while threats like shadow AI and adversarial attacks highlight the urgent need for robust governance.

Getting Started with AI Applications

Microsoft Guide for Securing the AI-Powered Enterprise

Read today

To help organizations navigate these challenges, Microsoft has released the Microsoft Guide for Securing the AI-Powered Enterprise Issue 1: Getting Started with AI Applications—the first in a series of deep dives into AI security, compliance, and governance. This guide lays the groundwork for securing the AI tools teams are already exploring and provides guidance on how to manage the risks associated with AI. It also dives into some unique risks with AI agents and how to manage these. Here’s a look at the key themes and takeaways.

Securing AI applications: Understanding the risks and how to address them

AI adoption is accelerating, bringing remarkable opportunities but also a growing set of security risks. As AI becomes more embedded in business decision-making, challenges such as data leakage, emerging cyber threats, and evolving and new regulations demand immediate attention. Let’s explore the top risks and how organizations can address them.

Data leakage and oversharing: Keeping AI from becoming a liability

AI thrives on data. But without guardrails, that dependence can introduce security challenges. One major concern is shadow AI—when employees use unapproved AI tools without oversight. It’s easy to see why this happens: teams eager to boost efficiency turn to freely available AI-powered chatbots or automation tools, often unaware of the security risks. In fact, 80% of business leaders worry that sensitive data could slip through the cracks due to unchecked AI use.²

Take a marketing team using an AI-powered content generator. If they connect it to unsecured sources, they might inadvertently expose proprietary strategies or customer data. Similarly, AI models often inherit the same permissions as their users, meaning an over-permissioned employee could unknowingly expose critical company data to an AI system. Without proper data lifecycle management, outdated or unnecessary data can linger in AI models, creating long-term security exposure.

Addressing the risk

Implement clear AI usage policies so employees stick to vetted, secure tools.
Allow AI’s to access data only based on the user’s credentials and use role-based access controls (RBAC) to manage that data access.
If an AI does need to use its own credentials to access resources, only grant AI access to people who should have access to those resources.
Automate data retention policies to prevent sensitive information from persisting beyond its intended use.

Emerging threats: The expanding landscape of AI vulnerabilities

As AI evolves, so do the threats against it. According to Gartner® Peer Community, among 332 participants, a staggering 88% of organizations are concerned about the rising risk of indirect prompt injection attacks,³ with attackers developing new ways to exploit vulnerabilities. One of the most pressing concerns is prompt injection attacks—where malicious actors embed hidden instructions in input data to manipulate AI behavior. A cleverly worded query, for example, could trick an AI-powered chatbot into revealing confidential information.

Beyond direct attacks, AI systems themselves can introduce security risks. AI models are prone to hallucinations (generating false or misleading information), unexpected preferences (amplifying unfair decision-making patterns), omissions (leaving out critical details), misinterpretation of data, and poor-quality or malicious input leading to flawed results. A hiring tool, for example, might favor certain candidates based on biased historical data rather than making fair, informed decisions.

Addressing the risk

Validate and sanitize input data before AI processes it to prevent manipulation.
Limit AI’s access to sensitive information and implement identity verification.
Design (and secure) the end-to-end (E2E) business process, not the AI component.
Secure the AI component by imagining replacing it with a new hire fresh out of school. What would you do to make your process robust against it making mistakes?
Use robust monitoring, validation, and oversight to ensure AI operates securely and responsibly.
Leverage commercial AI tools with built-in safeguards like bias detection, input sanitization, and access controls to reduce risks while supporting business goals.

Compliance challenges: Navigating the complex AI regulatory landscape

Beyond security, compliance is another major hurdle in AI adoption. Over half of business leaders (52%) admit they’re unsure how to navigate today’s rapidly evolving AI regulations.² Frameworks like the European Union AI Act, General Data Protection Regulation (GDPR), and Digital Operational Resilience Act (DORA) are rapidly evolving, making compliance a moving target. Organizations must establish clear governance and documentation to track AI usage, decision-making, and data handling, reducing the risk of non-compliance. Digital resilience laws like DORA require ongoing risk assessments to ensure operational continuity, while GDPR mandates transparency in AI-powered decisions like credit scoring and job screening. Misclassifying AI risk levels—such as underestimating the impact of a diagnostic AI tool—can lead to regulatory violations. Staying ahead requires structured risk assessments, automated compliance monitoring, and continuous policy adaptation to align with changing regulations.

Addressing the risk

Align AI governance with evolving regulations like DORA through risk assessment, testing, and monitoring.
Keep detailed AI records to ensure compliance and responsible use.
Use AI-powered tools to monitor compliance with GDPR and HIPAA, reducing risks like data drift and unauthorized access.
Audit AI decisions to ensure fairness, transparency, and regulatory compliance.
Classify AI risks clearly and stay updated on evolving regulations to maintain compliance.

The next frontier: Unique challenges in securing agentic AI

The pace of AI growth is staggering, with AI capabilities doubling every six months. Organizations are rapidly adopting more autonomous, adaptable, and deeply integrated systems to tackle complex challenges.

One of the most significant developments in this shift is agentic AI—a new class of AI systems designed to act independently, make real-time decisions, and collaborate with other AI agents to achieve complex objectives. These advancements have the potential to revolutionize industries, from optimizing energy grids to managing fleets of autonomous vehicles.

But with greater autonomy comes greater risk. Overreliance on AI outputs, cyber vulnerabilities, and reliability concerns all need to be addressed. As these systems integrate deeper into operations, strong security, oversight, and accountability will be essential.

Building a secure AI future: A responsible AI adoption playbook

AI’s transformative power comes with inherent risks, requiring a proactive, strategic approach to security. A Zero Trust framework ensures that every AI interaction is authenticated, authorized, and continuously monitored. But security isn’t something that happens overnight—it requires a phased approach.

Microsoft’s AI adoption guidance, part of the Cloud Adoption Framework for Azure, provides a structured path for organizations to follow and is clearly outlined in the Microsoft Guide for Securing the AI Powered Enterprise Issue 1: Getting Started with AI Applications. This guide offers a starting point for embracing the cultural shift needed to secure AI with clarity and confidence.

Cross-team collaboration, employee training, and transparent governance are just as essential as firewalls and encryption. By embedding security at every stage, breaking down silos, and fostering trust, organizations can confidently navigate the AI landscape, ensuring both innovation and resilience in a rapidly evolving world.