Your team may already be using artificial intelligence at work. Someone drafts emails with an online chatbot. Another employee summarises client documents. A third uses an AI tool to create reports or analyse spreadsheets.
Understanding the AI security risks involved is crucial for your team’s effective use of artificial intelligence.
That is shadow AI: artificial intelligence being used without formal approval, security checks or clear ownership.
The next step is more powerful. Instead of simply answering a question, AI can increasingly take action on your behalf. It can search systems, call applications, send messages, move files and complete multi-stage tasks with limited human input.
This is known as agentic AI.
The National Cyber Security Centre (NCSC) published guidance on managing the cyber risk of agentic AI on 20 August 2026. Its advice is practical and important for businesses of every size: sandbox agents, limit their access, monitor their activity and retain the ability to stop them immediately.
Addressing potential AI security risks can help in formulating a robust approach to technology management.
For a small business, these controls may sound like something only large technology companies need. They are not. If an AI agent can access your Microsoft 365 environment, customer records or financial systems, it must be secured like any other powerful user or business application.
What is agentic AI?
A standard chatbot responds to a prompt. You ask it to draft an email, and it gives you some text to review.
An AI agent can go further. You give it a goal, and it may decide which steps to take to achieve it. Depending on how it is configured, it could:
-
- Search your internal knowledge base.
When engaging with AI security risks, it is vital to be proactive in your measures.
- Read and organise documents.
- Update records in a customer relationship management system.
- Create and send emails.
- Book appointments.
- Use online services through application programming interfaces, or APIs.
- Trigger other software tools.
- Continue working without someone approving every individual action.
That autonomy is what makes agentic AI useful. It is also what creates new AI security risks for small business.
Fostering awareness around AI security risks is essential for minimising potential drawbacks.
An agent does not have human judgement or common sense. It interprets instructions based on its model, its available tools and the information it can access. If it misunderstands a task, encounters malicious content or receives a manipulated instruction, it may take an action that nobody intended.
Note: An AI agent is not automatically malicious. The risk comes from giving an autonomous system too much access, too little oversight or no reliable way to intervene.
Why should small businesses care about agentic AI?
Understanding the implications of AI security risks can enhance business operations.
Consider a fictional ten-person professional services firm in Winchester.
One employee introduces an AI agent to help manage enquiries. The agent can read incoming emails, search a shared folder for relevant documents and draft responses. To make the process faster, it is also given access to the company’s customer database and calendar.
At first, everything works well.
Then an email arrives containing hidden instructions designed to manipulate the agent. The email tells the agent to forward a confidential document to an external address and add a new rule to the mailbox. The agent follows the instruction because it appears relevant to the task it was given.
No traditional virus is required. The problem is excessive access combined with autonomous decision-making.
The potential impact could include:
- Unauthorised access to confidential information.
- Accidental disclosure of personal data.
- Fraudulent or misleading messages sent to customers.
- Unapproved changes to cloud systems.
- Deletion or alteration of business records.
- Compromised credentials being used across several services.
- Reputational damage and business interruption.
These are not reasons to avoid useful technology. They are reasons to introduce it properly.
Taking into account the AI security risks will lead to better decision-making processes.
Start with the lowest-risk use case
Before deploying an AI agent, decide whether you need autonomy at all.
Many tasks can be completed safely with a human reviewing the result. For example, an employee may ask an AI tool to draft an internal summary, but the employee remains responsible for checking and sending it.
Begin with tasks that have a limited impact if something goes wrong, such as:
- Drafting internal content.
- Summarising non-sensitive information.
- Classifying generic enquiries.
- Creating a first version of a report.
- Testing workflows with dummy data.
Avoid giving a new agent direct access to:
Effective strategies should mitigate known AI security risks for small businesses.
- Banking or payment systems.
- Payroll.
- Production databases.
- Customer personal data.
- Administrator accounts.
- Security settings.
- Unrestricted email sending.
- All company files or cloud applications.
Ask a simple question: What is the worst realistic outcome if this agent behaves unexpectedly?
If the answer involves a major financial loss, a data breach or prolonged downtime, the deployment requires stronger controls and human approval.
The nature of AI security risks requires careful consideration when deploying new technologies.
Apply the NCSC’s safeguards
The NCSC guidance sets out a useful framework for managing agentic AI. Here is how a small or medium-sized business can apply it.
1. Use a robust sandbox
A sandbox is a controlled environment that separates an application from the rest of your systems.
Always run an AI agent in a sandbox appropriate to the risk. Do not install an experimental agent directly on a staff member’s everyday laptop or place it inside the same unrestricted environment as your business data.
A properly designed sandbox should limit:
- The networks the agent can reach.
- The applications and tools it can use.
- The files and data it can access.
- The computing resources it can control.
- The credentials available to it.
- The systems it can influence indirectly.
Where possible, deny network traffic by default. Then allow only the specific services the agent needs. This is safer than allowing unrestricted internet access and hoping the agent behaves correctly.
Pro tip: Test an agent using sample data first. If it cannot complete the task without broad access to your real systems, that may be a sign that the workflow needs redesigning.
Realising the potential AI security risks is the first step in securing your systems.
2. Give the agent its own identity
Never allow an agent to operate using a shared administrator account or a member of staff’s personal credentials.
Create a distinct identity for each agent. This makes its activity easier to identify and allows you to disable that agent without disrupting every other user.
Use the principle of least privilege, which means giving an account only the permissions required for its task.
For example, an agent that drafts customer responses may need to read a specific mailbox. It should not automatically be able to:
- Delete mail.
- Change mailbox security settings.
- Access finance folders.
- Create new user accounts.
- Download the entire company file store.
Use short-lived credentials where possible. Revoke access when a task ends or when the agent is no longer required.
This is a core part of effective cybersecurity services in Winchester. Strong access controls reduce the damage if an account, application or agent is compromised.
3. Log and monitor activity continuously
Monitoring can expose any unexpected AI security risks that may arise during operation.
You cannot secure what you cannot see.
The NCSC recommends that agent activity is logged, audited and included in security operations. For a business using autonomous agents, this should include:
- The agent’s identity.
- The systems and tools it accessed.
- The instructions and inputs it received.
- The actions it attempted.
- The data it read or changed.
- Network connections it made.
- Errors, warnings and blocked actions.
- Human approvals and interventions.
Agentic AI activity should be treated as a form of user activity. That means monitoring it alongside human accounts and applications, ideally with 24/7 alerting and response.
Look for unusual behaviour, such as:
- A sudden increase in file access.
- Connections to an unapproved domain.
- Attempts to use administrator functions.
- Large numbers of external messages.
- Repeated attempts to bypass a restriction.
- Activity outside the agent’s normal working pattern.
Good logging also helps you investigate an incident. Without reliable records, it may be difficult to establish what happened, what information was involved and whether further action is needed.
4. Define red lines and human approval points
Write down what the agent must never do.
Defining the AI security risks that agents must avoid is critical for safe operation.
These red lines might include:
- Sending an external email without approval.
- Deleting or overwriting records.
- Accessing personal data unrelated to its task.
- Changing security controls.
- Creating new accounts.
- Making payments or changing bank details.
- Sharing confidential documents outside the business.
Then define when a person must approve an action. This is particularly important for high-impact tasks.
The NCSC describes different levels of oversight:
- Human-in-the-loop: a person approves actions before they happen.
- Human-on-the-loop: the agent acts, but a person monitors and can intervene.
- Human-out-of-the-loop: the agent acts without human review.
For a new deployment, start with human approval. You can increase autonomy later, once testing demonstrates that the controls work and the risks are understood.
Your strategy should address how to handle unexpected AI security risks effectively.
Do not rely on the agent’s prompt alone. Clear instructions are helpful, but prompts can be misunderstood, manipulated or forgotten during a long-running task. Use technical controls as well.
Keep an emergency “pull the plug” capability
The most important question is this: Can you stop the agent immediately?
The NCSC says organisations should always be able to halt autonomous activity if an incident is detected or reported.
Stopping one process may not be enough. Your emergency procedure may also need to:
- Disable the agent’s identity.
- Revoke its credentials and sessions.
- Cut its network access.
- Block communication with the AI model or inference service.
- Prevent connected tools from carrying out further actions.
- Preserve logs for investigation.
- Notify the person responsible for the deployment.
Name the people authorised to activate the shutdown. Document the steps and test them. A control that exists only on paper will not help during a fast-moving incident.
What should a Hampshire business do now?
A comprehensive plan should include responses to identified AI security risks.
If you run a small business in Winchester or elsewhere in Hampshire, take these practical steps:
-
- Create an AI register: record which tools and agents staff are using.
- Update your acceptable-use policy: explain which AI services are approved and what data must not be entered.
- Review Microsoft 365 permissions: remove unnecessary access and administrator rights.
- Separate sensitive information: do not allow experimental tools to search every shared folder.
- Test before deployment: use dummy data and a controlled environment.
By identifying AI security risks, you can create a more resilient business framework.
-
- Enable logging and alerts: make agent activity visible to your IT or security provider.
- Document shutdown procedures: include credential revocation and network isolation.
- Review suppliers: check how third-party AI services store, process and protect your information.
- Train employees: explain shadow AI, data handling and the risks of approving unexpected actions.
Training employees about AI security risks will empower them to use technology responsibly.
This is where IT security for small business in Hampshire can provide practical support. You do not need to build a security operations centre from scratch, but you do need clear ownership, sensible controls and ongoing monitoring.
A trusted provider offering managed IT services in Hampshire can help assess your environment, tighten access, monitor activity and support your response if something goes wrong.
Consulting with experts can help address AI security risks effectively.
Book a call about your AI security risks
Agentic AI could save your business time and help your team focus on higher-value work. But autonomy must be earned through testing, containment and oversight.
While leveraging AI, be vigilant regarding potential AI security risks that may emerge.
The NCSC’s 20 August 2026 guidance is a useful starting point. Sandbox the agent. Restrict its access. Monitor what it does. Keep the ability to stop it. Review those controls regularly as the technology and your business change.
If you are unsure whether your current AI use is safe, book a call with BITSmart Technology. We can discuss your systems in plain English and help you take practical steps towards safer adoption.
Understanding and managing AI security risks is essential for safe adoption.
