Guide AI Security

Agentic AI Security Risks: What Happens When AI Acts on Your Systems

A chatbot that answers questions is one security problem. An AI agent that can read your email, query your database and send messages on your behalf is a bigger one. Agentic AI moved from demos into production use in 2026, and many security programmes were not designed for software that acts on its own.

31 July 2026
6 min read
Updated 10 November 2026
Key takeaways
  • Agents move AI from reading to acting: they create, delete, send and trigger through tools. What an agent can do is set by the permissions and tools you give it.
  • Prompt injection is one of the main threats. Models cannot reliably separate trusted instructions from instructions hidden in emails, documents or web pages they process.
  • Over-permissioning is a recurring risk. Scope each agent to the resources its task needs and use short-lived credentials.
  • Enforce authorisation outside the model. Deterministic checks in the API or execution layer decide what is allowed, whatever the model asks for.
  • Shadow agents do not always leave an OAuth consent record. Discovery needs several sources: identity provider, low-code platform inventories, SaaS admin consoles and network logs.
  • Plan for a compromised agent: a kill switch must revoke tools and credentials, not just close the chat interface.

What agentic AI does

A chatbot returns text. An agent uses tools. Tools are integrations that let the AI take actions: call an API, read and write files, query a database, send email, run code, browse the web, trigger workflows or create calendar events. The model decides which tools to call, and in what order, based on the goal it has been given.

Common enterprise examples include Microsoft 365 Copilot agents that draft email, read SharePoint documents and update CRM records; models connected through MCP (Model Context Protocol) servers to internal knowledge bases and ticketing systems; customer service agents that look up orders and issue refunds; and IT agents that provision users, assign licences and reset passwords.

Most earlier AI deployments were read-only. They summarised documents, answered questions and generated text. Agents can be read-write: they create, delete, modify, send and trigger. That change is what alters the risk model.

Assess each agent before it goes live

Six questions give a quick first view of how much harm an agent could cause and how much control it needs. Answer them for each agent and record the answers in your agent inventory.

QuestionWhy it mattersHigher-risk answer
What can it read?Sets the data an attacker could extract through the agentMailboxes, whole file stores, customer or HR data
What can it change?Sets the damage a manipulated agent could doPayments, user accounts, production records, external messages
Who authorises its actions?Shows whether a person or a policy check sits between the model and the actionNo approval step and no independent policy check
Does it process external content?External content is the main route for indirect prompt injectionInbound email, public web pages, supplier documents, customer uploads
What happens if it fails or is manipulated?Sets the impact on customers, money and complianceFinancial loss, data breach, irreversible changes
Can it be stopped quickly?Sets how long a compromise can runNo kill switch, shared credentials, no owner on call

The new attack surface

74%
of IT application leaders surveyed by Gartner believe AI agents represent a new attack vector (Gartner, September 2025)
13%
strongly agree they have the right governance structures in place to manage AI agents (same survey of 360 leaders)

The attack surface has four layers. First, the model input: anything the model processes can carry adversarial instructions, and the model has no reliable way to tell them apart from legitimate ones. Second, the tool integrations: each API connection is an execution path, and every tool added to an agent widens what a successful attack can reach. Third, the data in the context window: documents, emails and database results the agent reads can contain embedded instructions that the agent then acts on. Fourth, the orchestration layer: in multi-agent systems, one agent delegates to another, so a manipulated sub-agent can take actions the primary agent was never meant to approve.

Speed changes the picture. An agent can run many tool calls in the time a person makes one, across every resource its permissions reach. A manipulated agent can therefore cause harm before a person notices, and the reach of that harm depends on the permissions and tools you have given it. This is why permission scoping and runtime limits matter more than detection alone.

Over-permissioning and blast radius

Least privilege applies to agents as it does to people and service accounts. Over-permissioning is a recurring risk because broad access is faster to set up than precise scopes, and the cost only shows when something goes wrong.

Look at how a permission is granted, not only its name. In Microsoft Graph, Files.ReadWrite.All granted as a delegated permission lets the agent act on files the signed-in user can reach. Granted as an application permission, it reaches files across the whole tenant without a user. For an agent that needs one SharePoint site, Microsoft's Selected permissions (Sites.Selected and the newer per-list and per-item variants) restrict application access to the resources you explicitly grant.

Common patterns to look for: tenant-wide file or mail permissions where one site or folder would do; agents that can call any external endpoint instead of an allowlist; and shared service accounts that have collected permissions from several projects.

To estimate an agent's blast radius, list every API, system and data store it can reach, and for each one describe what a manipulated agent could do with it at machine speed. If that outcome is unacceptable for the value of the task, narrow the permissions until it is acceptable. Monitoring helps you notice misuse; it does not shrink what the agent can reach.

Give each agent its own identity. Prefer managed identities or workload identities with short-lived credentials over long-lived secrets, so that revoking one agent does not affect others and audit trails stay clear. When an agent acts for a specific user, use delegated authorisation so it works within that user's rights, not with broader rights of its own. Our guide to machine identity security covers how to manage these identities.

Controls and governance in detail

Two companion articles cover the controls in detail.

Securing AI agents explains how direct, indirect and memory-based prompt injection work, why authorisation checks belong outside the model, which runtime limits contain a manipulated agent, how to review MCP servers and how to test agents before launch.

Governing AI agents covers finding shadow agents built on low-code platforms, keeping an agent inventory, matching human approval to the risk of each action, logging tool calls, and the steps to take when an agent is compromised.

Controls before your next agent deployment

  1. Assess the agent. Answer the six risk questions above and record the result in your agent inventory.
  2. Scope its permissions. Document every API, data store and system it will reach. Use the narrowest scope available, such as Selected permissions instead of tenant-wide access, and record why each permission is needed.
  3. Give it its own identity. Use a managed or workload identity with short-lived credentials, and delegated authorisation when it acts for a user. No personal credentials or shared accounts.
  4. Enforce authorisation outside the model. Put policy checks and parameter validation in the execution layer.
  5. Set runtime limits and a kill switch. Rate and transaction limits, an egress allowlist and a tested way to stop the agent and revoke its credentials.
  6. Add risk-based approval. Human approval for high-risk and critical actions, with the exact parameters shown to the approver.
  7. Turn on tool call logging. Capture identity, purpose, tool, resource, result, correlation ID and authorisation decision, with minimisation and access control.
  8. Test, then review. Test in proportion to risk before launch, review on every change, and run your shadow agent discovery on a regular cycle.
Ryland Deakin
About the author
Lead Consultant, Cyvra · CISM · CompTIA Security+ · MCP

Ryland has delivered cybersecurity, compliance, and IT management programmes for regulated organisations across the UK and the Netherlands for over 20 years. View full profile

Common questions

How is an AI agent different from a standard API integration?

A standard integration runs deterministic code: if condition A, take action B. An agent decides what to do based on natural language context. That makes it flexible but less predictable. It may read an ambiguous instruction differently from what the developer intended, or follow adversarial content that a deterministic integration would ignore. This is why agents need authorisation checks outside the model and testing that goes beyond the expected path.

Sources

AI Security Assessment

Understand what your AI agents can do

We build an inventory of your AI agents, assess each one by risk, review permissions, tools and identities, and give you prioritised recommendations with a governance approach that fits your environment.