Artificial intelligence is changing how cyber attacks are carried out. But as businesses introduce their own AI tools and agents, are they creating risks they don't fully understand?
Artificial intelligence is becoming part of everyday business operations. Employees are using AI to draft emails, summarise meetings, analyse documents and support decision-making. Businesses are also exploring AI agents that can connect to systems, retrieve information and carry out tasks with limited human intervention. These technologies offer opportunities to improve efficiency, revenue and reduce repetitive work.
But AI is also changing the other areas including cyber threat’s.
In November 2026, the UK's National Cyber Security Centre (NCSC) published its Annual Review, warning that AI is increasing the speed, scale and efficiency of cyber attacks. For businesses, this raises two important considerations.
First, how might attackers use AI to target your organisation? Second, what new risks are you introducing through the AI tools and agents your own business uses?
Both deserve attention, particularly as organisations increasingly connect AI systems to customer information, internal applications and business processes.
What is changing about AI-enabled cyber attacks?
Cyber criminals have traditionally relied on techniques such as phishing, stolen credentials, malicious software and exploiting weaknesses in business systems. These methods have not disappeared. What is changing is how quickly and efficiently attackers may be able to carry them out.
According to the NCSC's 2026 Annual Review, AI is already being used to support several stages of cyber operations, including reconnaissance, social engineering, vulnerability discovery and exploit development. In practical terms, AI can make it easier for attackers to research potential targets, create convincing messages, identify weaknesses and automate repetitive activities. An attacker who previously needed considerable time to research an organisation may now be able to use AI to accelerate parts of that work. However, it is important not to overstate what AI can currently do.
The NCSC assesses that human operators remain important in directing and adapting cyber operations. Fully autonomous, end-to-end AI attacks are not yet the predominant approach. The immediate concern is that AI is making existing attack methods faster and more scalable.
What are AI agents, and why do they matter?
An AI agent is a system that can pursue a defined objective by planning and carrying out actions, often using connected tools or applications. Unlike a conventional chatbot that primarily responds to questions, an AI agent may be able to take multiple steps towards completing a task. For example, an organisation might introduce an AI agent to support customer enquiries. Depending on its configuration and permissions, the agent could:
Retrieve information from a customer database.
Check the status of an order or booking.
Update a customer record.
Create a support ticket.
Send a message to a customer via different communication tools like email and WhatsApp.
Escalate an issue to an employee.
These capabilities can be useful, but they also introduce new considerations because the AI system may have permission to access information or perform actions that affect the business and its customers. The more systems an AI agent can access, and the more independently it can operate, the greater the need for appropriate oversight.
How are attackers using AI agents?
AI agents can support attackers by automating parts of their activities. For example, AI-enabled tools may be used to research an organisation, identify exposed systems, analyse software weaknesses or prepare targeted phishing messages.
Some agentic systems can coordinate multiple steps rather than relying on a human to issue every individual instruction. This could allow attackers to attempt certain activities more efficiently. However, the existence of an AI agent does not mean it can successfully compromise a business without human involvement. AI systems can make mistakes, misinterpret information and fail to complete tasks.
The NCSC's assessment is that AI is currently more likely to accelerate existing attack methods than replace human attackers entirely. For businesses, the practical implication is that familiar weaknesses remain important. An account without appropriate authentication, an unpatched application or excessive user permissions can still provide an opportunity for an attacker. AI may simply make it easier or faster for an attacker to identify and exploit those weaknesses.
AI also creates risks inside your own business
While attention often focuses on how cyber criminals use AI, organisations should also consider the risks associated with their own AI adoption. Many businesses are introducing AI tools without fully understanding what information those tools can access, where that information is processed or what actions the systems can perform. This is particularly relevant when employees connect AI applications to existing business systems.
Consider a business using an AI meeting assistant. The tool may request access to calendars, meeting recordings, transcripts and connected documents. Or consider an AI agent introduced to support customer service. It may require access to customer records, email systems and internal knowledge bases. These integrations can be useful, but they also expand the number of applications and services that can access business information.
The question is no longer simply whether an AI tool is useful. It's what the tool can access, what it can do and who is accountable for its actions.
Shadow AI: When employees introduce tools without approval
One of the concerns highlighted in the NCSC's 2026 Annual Review is shadow AI. Shadow AI occurs when employees use AI applications or services without appropriate organisational approval or oversight. For example, an employee might upload a customer spreadsheet to an AI tool to generate a report. Another might connect an AI meeting assistant to their company calendar. A contractor might use an AI application to summarise confidential project documents. These activities may be intended to improve productivity. However, the organisation may not know:
Which AI tools are being used.
What personal or confidential information is being shared.
Where the information is processed or stored.
Whether the supplier uses submitted information to improve its models.
What access permissions have been granted.
Whether information can be deleted or access revoked.
This creates potential security, privacy and contractual risks.
A business may have established controls over its internal systems but lose visibility when employees transfer information to external AI services.
The NCSC has specifically warned that shadow AI can increase organisational vulnerability by reducing visibility over how information is handled. For businesses, the solution is not necessarily to prohibit every AI tool. It is to establish clear expectations about which tools may be used, what information may be shared and how new AI applications are approved.
What happens when an AI agent has too much access?
Consider a business that introduces an AI agent to support its customer service team. The agent is connected to the organisation's customer database, email platform and document storage. Its intended purpose is to retrieve information and support staff in responding to customer enquiries. However, the agent has been granted broad permissions to access records and modify information. What happens if the agent misunderstands an instruction, retrieves the wrong customer's information or performs an action that was never intended? What if an attacker places malicious instructions inside a document or message that the agent is asked to process? This is one of the risks associated with prompt injection.
Prompt injection occurs when an attacker attempts to influence an AI system's behaviour through instructions embedded in content the system encounters. For example, a malicious instruction hidden in an email or document could attempt to make an AI agent ignore its intended task, disclose information or misuse a connected tool. Whether the attempt succeeds depends on the system's design, permissions and safeguards. But it illustrates why AI agents require more than a general instruction to behave responsibly.
Businesses should consider technical restrictions, access permissions, testing, monitoring and human approval for higher-risk actions. An AI agent should not automatically receive unrestricted access simply because doing so makes a workflow more convenient and generate more revenue.
Why AI governance matters
AI governance is about establishing how AI is selected, approved, used, monitored and held accountable within an organisation. It helps businesses answer practical questions such as:
Who is responsible for approving new AI tools?
What business purpose does each AI system serve?
What information can the system access or process?
Has the AI supplier been assessed?
What decisions or actions can the AI system perform?
When is human review required?
How are errors, incidents and complaints handled?
How is the system monitored after deployment?
These responsibilities become particularly important when AI systems can act on behalf of employees or influence decisions affecting customers. For example, a business might use AI to recommend financial products, prioritise customer complaints or evaluate applications. In these situations, the organisation needs to understand how the AI contributes to decisions and whether the use of personal information is appropriate.
Governance is not simply about producing an AI policy or deploying a one hour power point presentation to employees each year. It involves ensuring that responsibilities, processes and safeguards are implemented in everyday operations.
Privacy risks should not be overlooked
AI systems frequently rely on information to generate responses or carry out tasks. That information may include personal data, confidential business records or commercially sensitive material.
Under the UK GDPR, organisations processing personal data must have an appropriate lawful basis, provide relevant transparency information and implement suitable security measures.
Depending on the activity, organisations may also need to undertake a Data Protection Impact Assessment (DPIA). For example, before connecting an AI agent to customer records, a business should consider:
Purpose: Why does the AI system need access to the information?
Data minimisation: Does the agent require access to every customer record, or only the information necessary for a specific task?
Supplier arrangements: Who processes the information, where is it processed and what contractual protections apply?
Access controls: Can the AI system retrieve, modify, export or delete information?
Retention: How long is information retained by the AI provider?
Human oversight: What happens if the system produces an incorrect response or takes an inappropriate action?
These are not simply technical questions. They are business governance and privacy decisions that require clear ownership from business leaders, boards, and business owners.
Does the UK Cyber Essentials protect businesses against AI-related risks?
UK Cyber Essentials provides an important foundation for managing common cyber security threats. Its five technical controls address areas such as secure configuration, software updates, malware protection and user access management. These remain relevant as attackers adopt AI-enabled techniques. For example, an organisation with properly managed access controls and up-to-date systems may be better positioned to resist certain attacks, regardless of whether AI was used to support them.
However, Cyber Essentials is not an AI governance certification.
It does not, by itself, establish that an organisation has:
Approved and assessed its AI tools.
Identified all AI systems in use.
Evaluated the privacy implications of AI processing.
Established accountability for AI-generated decisions.
Restricted the actions AI agents can perform.
Implemented appropriate human oversight.
Assessed AI-specific risks such as prompt injection.
Businesses therefore need to consider how their existing security arrangements connect with broader AI governance and privacy responsibilities.
The UK Government has recently introduced a 75% discount on Cyber Essentials certification for eligible small and medium-sized businesses. We explore the funding, eligibility and practical considerations in our related article: UK Cyber Essentials: 75% Government Discount – Is Certification Worth It for Your Business?
Seven practical steps businesses can take
Businesses do not necessarily need a large AI governance programme to begin addressing these risks. A practical starting point is to understand where AI is being used and introduce proportionate controls.
1. Identify the AI tools already in use
Create a register of AI applications used across the business, including those introduced by employees and contractors. Record their purpose, supplier, users and the information they can access.
2. Establish an AI approval process
Before introducing a new AI application, assess its business purpose, privacy implications, security arrangements and supplier practices. Not every tool requires an extensive assessment, but higher-risk applications should receive closer scrutiny.
3. Review AI access permissions
Understand which systems AI tools and AI agents are connected to. Limit access to the information and functions necessary for their intended purpose. Where possible, avoid granting broad administrator permissions or unrestricted access to customer records.
4. Establish clear rules for handling information
Define what employees may and may not share with AI tools. This should include guidance on personal information, confidential documents, customer records and commercially sensitive information.
5. Introduce human oversight for important actions
AI agents should not automatically be permitted to perform every action independently regardless what you may hear on social media or those that promote it on YouTube. Consider requiring human approval before sending sensitive communications, modifying important records, transferring information or carrying out actions with significant consequences.
6. Monitor AI activity and manage incidents
Where appropriate, maintain records of AI actions, monitor unusual behaviour and establish a process for reporting problems. Your incident response arrangements should consider incidents involving AI systems, including unauthorised information disclosure or inappropriate automated actions.
7. Review AI use as the business changes
AI governance should not be a one-time assessment either as it involves ongoing monitoring in the same way as employee undergoes performance reviews and regular feedback on tasks progress. As AI tools gain new features, permissions or integrations, their risks may change. Review them periodically and when significant changes occur and not by an AI agent, but by a human.
The bigger issue: Who is accountable when AI acts?
One of the most important questions businesses need to address is accountability.
If an employee makes an error, an organisation will generally have established processes for reviewing what happened and deciding how to respond. But what happens when an AI agent makes the error?
Who approved the agent? Who determined what information it could access? Who authorised the actions it could perform? Who monitors whether it is operating as intended?
The organisation cannot assume that responsibility transfers to the AI system simply because an action was automated. The allocation of legal responsibility will depend on the circumstances, but businesses remain responsible for meeting their own applicable obligations. This is why AI governance needs to be considered before an AI system is connected to important business processes, not only after something goes wrong.
AI is changing the risks, but the fundamentals still matter
The NCSC's 2026 findings reinforce an important point. AI is changing the speed and scale of cyber threats, but many of the underlying vulnerabilities remain familiar. Businesses still need secure systems, appropriate access controls, supported software and effective incident response arrangements. At the same time, their own adoption of AI introduces additional questions about privacy, permissions, supplier oversight and accountability. Good cyber security provides a foundation. AI governance establishes how AI can be used responsibly within that foundation. Privacy ensures personal information is handled appropriately. These areas should work together rather than being managed in isolation.
For organisations, the starting point is not necessarily purchasing another tool or introducing a complicated framework or activating another AI agent. It is understanding what AI is already doing within the organisation, what information it can access and what safeguards are needed.
How Hyplon supports businesses
At Hyplon, we support organisations in establishing practical AI governance and privacy arrangements that align with how their businesses operate. This includes reviewing AI tools and suppliers, creating strategy roadmaps, appropriate trainging programs, reviewing privacy policies, assessing privacy risks, establishing approval processes, defining responsibilities and introducing appropriate oversight of AI systems and agents.
Our focus is on supporting businesses to adopt AI with greater confidence while maintaining accountability, protecting information, managing risk and preventing loosing customers and maintaining trust and reputation.
Before giving an AI agent access to your business systems, make sure you understand what it can do, what information it can access and who is responsible for its actions.
Useful resources
This article provides general guidance. The appropriate controls and legal obligations will depend on an organisation's AI use, information processing and operating environment.
