AI Agent Security: Protecting Data and Autonomous Workflows

Table of Contents

Dynamics 365 Activate: AI-Powered ERP Migration Explained
Dynamics 365 Activate: How AI Is Changing ERP Migration and ModernisationERP migrations have a reputation...
AI-Ready Infrastructure: What Businesses Need Beyond Models
Building AI-Ready Infrastructure: What Businesses Need Beyond AI ModelsChoosing an AI model is now...
AI in Procurement: Source-to-Pay with Dynamics 365
AI in Procurement: How Dynamics 365 Is Bringing Source-to-Pay Into the Tools People Already UseProcurement...
Microsoft IQ Explained: Business Context for AI Agents
Microsoft IQ Explained: Why AI Agents Need Business Context to Be UsefulAn AI agent that does not...
Microsoft Purview and Copilot: Data Security in the AI Era
Microsoft Purview and Copilot: Why Data Security Now Decides How Fast You Adopt AICopilot is only...
AI Agent Security: Protecting Data and Autonomous Workflows
AI Agent Security: How to Protect Business Data When Agents Start Acting on Their OwnAI agents no...

Leave a Reply

Your email address will not be published. Required fields are marked *

AI agent security dashboard showing agent risk levels across an organisation

AI Agent Security: How to Protect Business Data When Agents Start Acting on Their Own

AI agents no longer just answer questions. They read files, call tools, update records, and trigger workflows. As a result, AI agent security is now one of the most pressing questions for any organisation adopting AI this year.

The wider threat picture is already demanding. According to the Cyber Security Breaches Survey 2025/26, 43% of UK businesses reported a breach or attack last year. Therefore, adding autonomous agents to that environment, without the right controls, widens the attack surface.

This blog is for CEOs, CISOs, CIOs, and boards who are accountable for how AI uses company data. We explain what changed at Microsoft Ignite 2026 and what leaders should ask their teams now.

What This Means for Leadership

  • Accountability: boards will be asked how many agents they run and what those agents can reach.
  • Regulatory exposure: under UK GDPR, an agent that leaks personal data is still your organisation’s breach.
  • Governance as an enabler: clear controls let you scale AI with confidence instead of pausing it.
  • Three questions to ask: Do we have a full agent inventory? Does each agent have its own identity? Who approves what agents can access?

[Image alt text: “AI agent security dashboard showing agent risk levels across an organisation”]

Why Agents Create a New Kind of Risk

Traditional security assumes a person sits behind every action. However, agents break that assumption. An agent can hold permissions, connect to other systems through MCP servers, and act at machine speed.

Above all, the bigger problem is visibility. Staff install local agents on laptops, teams build agents in different tools, and plugins arrive without review. Consequently, many organisations cannot answer a simple question: how many agents do we run, and what can they reach?

Microsoft now calls agents “the new insiders”. Like any insider, organisations need to discover, onboard, and contain them. Meanwhile, the same Microsoft post reports more than 96,000 vulnerabilities published in 2026. For this reason, AI agent security now starts with discovery.

The Microsoft Approach to AI Agent Security

At Ignite 2026, Microsoft positioned Agent 365 as its observability platform for AI. It covers management, governance, security, and business value for agents in one place. Moreover, several new capabilities target the visibility gap directly.

Agent 365 has been generally available since May 2026. At Ignite, Microsoft then set out its next step: moving from an agent control plane to an AI observability platform. In practice, that means one view of agents, models, and assistants, with security, governance, and cost visibility side by side.

First, Dynamic Agent Risk, in preview, combines runtime signals, security posture, and potential blast radius into one assessment. Specifically, it draws on identity, access, data, endpoint, cloud, and network signals. Partners including Check Point, Cyera, Darktrace, and Zenity contribute signals through Microsoft Sentinel. Microsoft then calculates a single low, medium, or high risk level for each agent.

Second, governance now reaches further. Plugin and skill governance is generally available, while MCP server and tool governance is in preview. Meanwhile, local and shadow agent governance, also in preview, brings unsanctioned agents into view. In addition, an action centre is due in private preview in December.

Discovery has also widened. Agent 365 now discovers 60 local agent types, up from 25, including Claude Code, OpenAI Codex, and GitHub Copilot. It also reaches third-party cloud agents on AWS and Google Cloud, in preview. Furthermore, each discovered agent receives a Microsoft Entra Agent ID, so teams can govern it with Conditional Access.

Third, Microsoft Defender adds runtime protection for local agents. Agent-native protection is generally available for tools such as GitHub Copilot CLI and Claude Code. Similarly, network-based protection covers others, including ChatGPT and Ollama Desktop. Notably, this protection works alongside Microsoft Purview, Entra, and network data loss prevention.

Identity controls are getting sharper too. Entra ID Governance dynamic permissions, in limited preview, give agents task-scoped and time-bound access. Additionally, Intune policies can place supported local agents into managed execution environments on company devices.

Finally, Microsoft Perception is now generally available for security operations. It coordinates AI agents that probe for weaknesses, defend against attacks, and drive fixes through playbooks. Importantly, customers still approve high-impact actions, so people stay in control.

Key AI Agent Security Capabilities

CapabilityWhat it doesWho benefits
Dynamic Agent Risk (preview)Scores each agent as low, medium, or high based on live signals and blast radiusSecurity operations
Plugin and skill governance (GA)Controls which plugins and skills agents may useIT administrators
MCP server and tool governance (preview)Governs the tools and connections agents reachSecurity architects
Local and shadow agent governance (preview)Finds and governs agents running outside approved channelsCISOs
Agent usage insights (GA)Shows agent use in the Microsoft admin centreIT and business leaders
Defender runtime protection for local agents (GA)Protects developer and desktop agents at runtimeSecurity teams
Expanded agent discovery (preview)Finds 60 local agent types plus cloud agents on AWS and Google CloudIT administrators
Entra Agent ID for discovered agents (preview)Gives each agent an identity governed by Conditional AccessIdentity teams
Dynamic permissions (limited preview)Grants agents task-scoped, time-bound access onlySecurity architects
Microsoft Perception (GA)Runs security agents through playbooks, with people approving high-impact actionsSecurity operations

AI Agent Security in Practice by Sector

Healthcare and NHS

Trusts handle some of the most sensitive data in the country. For instance, an agent summarising referral letters must never pass patient data to an unapproved tool. To that end, Agent 365 governance lets a trust approve specific tools and block the rest. We bring NHS DSPT Standards Met status for 2025 to 2026 to these projects.

Public Sector

Councils face strict accountability for automated decisions. Therefore, a clear agent inventory and risk score support audit and scrutiny. We help public bodies map agents to policy before they scale.

Non-Profit

The same survey found that 28% of charities identified a breach or attack. Equally, charities hold donor and beneficiary data, often with small IT teams. Additionally, shadow agents can appear quickly when volunteers adopt new tools. We help non-profits set proportionate controls that do not slow their work.

Business Outcomes of Strong AI Agent Security

  • Faster detection of risky or misconfigured agents through unified risk scoring
  • Improved visibility of shadow agents running on staff devices
  • Reduced risk of sensitive data leaving through unapproved tools and plugins
  • Greater confidence for boards and auditors that agent use is governed
  • Least-privilege access, with agent permissions limited to the task and time needed

AI Agent Security FAQs

What is Microsoft Agent 365?

Microsoft Agent 365 is Microsoft’s platform for managing, governing, and securing AI agents across an organisation. It has been generally available since May 2026. At Ignite 2026, Microsoft extended it towards a wider AI observability platform. As a result, leaders get one view of agents, models, and assistants, with security, governance, and cost side by side.

How do we find shadow AI agents in our organisation?

Start with discovery. In preview, Agent 365 now recognises 60 local agent types, such as Claude Code and GitHub Copilot, plus agents on AWS and Google Cloud. Each discovered agent then receives a Microsoft Entra Agent ID. Consequently, your security team can apply the same access policies it already uses for staff.

Who is accountable if an AI agent leaks personal data?

Your organisation is. Under UK GDPR, the organisation that controls the data remains responsible, whether a person or an agent caused the leak. Therefore, boards should treat agents like any other identity. That means a full inventory, clear ownership, and access limited to what each agent needs.

Do we need to pause AI adoption to fix AI agent security?

No. In fact, good governance usually speeds adoption up. Clear controls give teams approved tools and give leaders the confidence to scale. We typically start with an agent inventory and a risk review, then put controls in place in phases.

Building Agent Security Before Agents Scale

Undoubtedly, agent numbers will keep climbing as adoption grows. That said, governance does not need to slow adoption. The organisations that succeed will treat agents like any other identity: inventoried, scored, and controlled.

As a Microsoft Solutions Partner delivering cloud, security, and managed services, we help organisations put these controls in place early. In turn, our team works across Agent 365, Defender, Sentinel, and Purview, so security and data protection move together.

At Mazik Global UK, we help organisations unlock the full potential of Microsoft technologies with AI-driven innovation and data-led strategies.

If you are ready to scale AI agents without losing control of your data, our team is here to help. Get in touch at mazikglobal.co.uk/contact-us.

Get Started

Let’s Talk About Your Next Step

Book a discovery call or send us a message.
We’ll get back to you within one business day.

A 30-minute call with our specialists to explore where AI and automation can make a real difference. There’s no sales pitch, just an honest conversation about your goals and challenges.

Schedule a discovery call

Tell us about your goals, and we’ll tailor our expertise to fit your needs. Fill out the form below, and we’ll get back to you soon.

We'll only use your info to respond to your inquiry.

Latest from Our Experts

Dynamics 365 Activate: AI-Powered ERP Migration Explained
20Nov

Dynamics 365 Activate: AI-Powered ERP Migration Explained

Dynamics 365 Activate: How AI Is Changing ERP Migration and Modernisation ERP migrations have a reputation for running long and…

AI-Ready Infrastructure: What Businesses Need Beyond Models
20Nov

AI-Ready Infrastructure: What Businesses Need Beyond Models

Building AI-Ready Infrastructure: What Businesses Need Beyond AI Models Choosing an AI model is now the easy part. The harder…

AI in Procurement: Source-to-Pay with Dynamics 365
20Nov

AI in Procurement: Source-to-Pay with Dynamics 365

AI in Procurement: How Dynamics 365 Is Bringing Source-to-Pay Into the Tools People Already Use Procurement teams spend much of…

Mazik Global UK, a trusted Microsoft Solution Partner and FastTrack-recognised expert, delivers AI-powered, low-code solutions across Dynamics 365, Power Platform, and Azure, driving confident digital transformation.

Subscription Form
Calendar event not found
Edit Template
Summit Insight report - Form
NFP Webinaer 26
On-Demand Webinar
NHS Webinar 2026