AI-ENABLED CYBER THREATS
A familiar voice. A convincing message. An urgent request. AI can make deception harder to spot—and help attackers move faster. Build security around what a request can do, not how genuine it appears.
APPEARANCE ≠ IDENTITY
Verify the request. Protect the decision.
THE CHANGE THAT MATTERS
AI-enabled threats use artificial intelligence to assist an attack: researching a target, generating persuasive content, imitating an identity or supporting technical intrusion. AI is an accelerator; it does not make every attack successful.
Your business can be targeted even if it does not use AI. The priority is to make consequential actions harder to trigger through deception, and to limit the damage when an account or device is compromised.
The UK NCSC assesses that AI will increase the effectiveness and efficiency of elements of cyber intrusion. Read the NCSC assessment.
01 / THE MESSAGE
A fraudulent message may reference a real colleague, a current project or an actual supplier. AI-generated text can remove awkward phrasing and help tailor the same approach to different people. Email, messaging apps and social platforms can all carry the request.
The key question is what happens next: an employee enters a password, approves a sign-in, downloads a file or changes a business record. Review those actions, not just the wording of the message.
A voice can be reproduced. Authority must be verified.
02 / THE IDENTITY
Voice cloning and synthetic video can add apparent credibility to a fraudulent request. A call that sounds like an executive, a supplier or a family member can create pressure to make an exception.
The FBI warns that criminals use AI-generated audio and video to impersonate people and support fraud. Visual or audio oddities may be a clue, but their absence does not establish authenticity. Read the FBI advisory.
End the unexpected interaction and contact the person using a number or process already held in trusted records. For payment changes, involve a separate approver and follow the established supplier-verification process. A new number supplied by the requester is not an independent channel.
THE PRINCIPLE
03 / THE ATTACK PATH
This simplified example follows a social-engineering intrusion. Real incidents vary, and a payment fraud may never involve malware at all. Your controls should interrupt the path at more than one point.
01
Public information helps make the approach relevant.
Reduce unnecessary exposure of roles, routines and sensitive project details.
02
A believable request creates urgency or apparent authority.
Check the business purpose and verify through a trusted, separate route.
03
A login, approval or vulnerable system creates a foothold.
Use phishing-resistant MFA where supported; patch exposed systems and limit privileges.
04
The attacker may reach more accounts, data or services.
Monitor activity, restrict lateral access and test containment and recovery procedures.
Do not wait for proof that AI was used. Investigate the suspicious behaviour and the access or action it enabled.
04 / BEYOND SOCIAL ENGINEERING
AI can support vulnerability research and other parts of technical attack preparation. The NCSC highlights the growing pressure on systems that have not received available security fixes. This makes asset ownership and patch decisions especially important.
Know which services are reachable from the internet, who maintains them and which vulnerabilities need urgent attention. Where immediate patching is not possible, agree a temporary mitigation and a clear review date.
An AI-assisted approach can lead into familiar incident outcomes: stolen credentials, disrupted operations, data theft or ransomware. Backups alone do not address stolen information or prevent a compromised identity from being reused.
Plan which services must return first, how recovery access is protected, and how restored systems will be checked before reconnection. Exercises should involve business owners as well as the technical team.
05 / YOUR OWN AI SYSTEMS
AI-enabled attacks and attacks against AI systems are different, but they can meet inside a business workflow. An assistant may read untrusted material and also have access to company information or tools.
Prompt injection attempts to influence the model through instructions embedded in input, such as a document or webpage. The potential impact depends on the access and actions available to the application. See OWASP’s prompt injection guidance.
Restrict permissions, separate untrusted content from authorisation, and require approval for consequential actions. Test the whole workflow, including connected systems.
ILLUSTRATIVE TRUST BOUNDARY
Information to review—not permission to act.
VERIFY WHAT CAN CROSS
Untrusted content must not authorise tools.
Limited data access. Defined tools. Human approval for sensitive actions.
MAKE IT RELEVANT TO YOUR TEAM
Open the area closest to your work. The useful conversation starts with the decisions your team makes every day.
A supplier asks to change bank details. A senior colleague wants an urgent transfer.
Agree how payment changes are verified, who provides independent approval and what staff should do when a senior person asks them to bypass the process. Keep trusted contact records separate from incoming messages.
Someone requests an MFA reset or emergency administrator access.
Check how identity is verified before a reset. Review privileged accounts, recovery routes and the records kept when access is restored. An apparently familiar voice should not replace the documented verification process.
A candidate, employee or customer sends an unexpected file or requests sensitive information.
Use approved document-handling and identity-checking processes. Define what information can be disclosed and make it easy to escalate a request that does not fit the normal workflow.
An unusual request could interrupt service delivery or change a critical system.
Identify the approvals required for operational changes and the people who can pause an action. Include suppliers, dependencies and recovery priorities in the response plan.
SOMETHING ALREADY FEELS WRONG?
If you suspect a fraudulent request, stop the proposed action and alert your internal security or IT contact. Preserve the original messages and useful timestamps. If money has moved, contact your bank promptly through a verified channel.
For a suspected account or device compromise, ask your response team to assess containment. Use a trusted device or separate communication route if the usual one may be affected.
Share a brief summary and a safe callback number. Keep passwords and sensitive evidence out of the enquiry form.
We will confirm availability and scope directly.
FROM AWARENESS TO ACTION
Bring your current IT team into the conversation. We can discuss a focused scope around the decisions, accounts and services that matter to your business.
01
Map sensitive requests: payments, password resets, data disclosures and operational changes. Agree the business owner for each.
02
Review the relevant settings, approval records and response instructions. Separate controls that are documented from controls that have been demonstrated.
03
Agree realistic improvements, responsible people and a way to validate progress. Discuss any deeper technical testing separately before it begins.
QUESTIONS, ANSWERED
AI assists one or more parts of the activity, such as generating a message, imitating a voice or supporting research. It does not need to control the entire attack. The response should focus on the behaviour, affected accounts and business impact.
Yes. An attacker can use AI independently of your technology choices. Staff, suppliers and public-facing services may still be approached. Your own AI adoption introduces a separate set of data, permission and workflow questions.
There is no dependable visual or listening test that establishes identity in every case. Training can encourage caution, but sensitive decisions should rely on independent verification and established approvals. Detection tools can contribute signals; they should not be the sole basis for trust.
MFA is an important layer, especially phishing-resistant methods where available. It does not prevent every form of payment fraud, session theft, deceptive approval or account-recovery abuse. Review sign-in, recovery and business approvals together.
This page focuses on hostile activity and the business controls that interrupt it. Secure AI adoption focuses on the tools your own team uses, the data they access and the actions they can perform. A useful programme connects both areas.
Start by understanding your current coverage and gaps. A change to verification, access or ownership may be more useful than another product. Any tooling recommendation should follow the agreed scope and the evidence collected.
Bring a summary of your business, your main concerns, relevant systems and who manages them. Describe any suspicious event at a high level. Do not send passwords, confidential customer files or incident evidence through a general enquiry form.
We can discuss an engagement that includes your existing team or service provider. Agree responsibilities, access, deliverables and communication routes at the outset so that improvement work has a clear owner.
KEEP BUILDING YOUR UNDERSTANDING
Connect these threats to the way your business uses data, makes decisions and prepares for disruption.
Further reading: NCSC: AI and the cyber threat · FBI: AI-enabled fraud · OWASP: Prompt injection. Scenarios and artwork on this page are illustrative.
Practical reading to help your team turn AI security questions into informed decisions.
A practical verification process for finance teams facing unexpected payment changes, executive impersonation and convincing voice calls.
Read article : A familiar voice is not approval: checking urgent payment requestsMove beyond spelling mistakes and examine the sign-ins, file handling and approvals that a persuasive message tries to trigger.
Read article : AI phishing: protect the action behind the messageUnderstand the difference between reading untrusted content and allowing that content to influence a connected business action.
Read article : Prompt injection: what changes when an AI assistant can act?MAKE TRUST A VERIFIED DECISION
Let’s identify where a convincing request could become a costly action—and what your business can put in its path.
A short form. A focused next step.
