Before You Click ‘Allow’: Cyber Security in the Age of AI

Keep these four questions in mind as you read:

• App permissions: does this request match the task?
• Account sign-in: can I use a passkey?
• AI agents: what can this system do without approval?
• AI-generated code: what needs reviewing before real people use it?

Consent phishing: the risk behind a genuine permission screen

Purpose: Check whether an application should have access to your account.

Consent phishing tricks someone into authorising a malicious application. Instead of collecting your password, the attacker gains the access you approve, which may include reading emails or files.

A September 2026 warning from the US Federal Bureau of Investigation (FBI) described attacks involving supposed file-sharing requests and event invitations. The permission screen can belong to a legitimate service while the app requesting access is malicious. Read the FBI’s consent phishing warning.

Verify unexpected requests through a separate, trusted channel. Read the requested permissions and stop if they do not fit the task. Familiar branding is not enough.

If you have approved a suspicious app, revoke its access in your account’s security settings. A password change alone may leave that access in place. For a work account, contact your IT or security team promptly.

Passkeys: a stronger way to sign in

Purpose: Reduce the risk of someone stealing and reusing your login details.

A passkey is a sign-in credential managed by your device or credential manager, usually unlocked with your fingerprint, face or device PIN. It uses cryptographic keys rather than a password you type into a website.

Passkeys resist phishing because authentication is tied to the correct service. A lookalike website cannot collect and reuse them as it could a password.

In April 2026, the UK’s National Cyber Security Centre (NCSC) began recommending passkeys wherever services support them. Read the NCSC’s assessment of passkeys.

Check one important personal account for passkey support and understand its recovery options. Where passkeys are unavailable, use a unique password stored in a password manager and enable two-step verification, which adds a check beyond your password.

AI agents: useful tools need clear limits

Purpose: Match an AI system’s access and independence to its task.

Agentic AI describes systems that pursue tasks through multiple steps, often using connected tools. An assistant might search documents, update a record and send a message as part of one workflow.

In May 2026, the US Cybersecurity and Infrastructure Security Agency (CISA) and international partners, including the NCSC, published guidance recommending limits on agent autonomy and access, alongside monitoring and oversight. Read the joint guidance announcement.

Apply least privilege: give the tool only the access necessary for its task. Check what it can read, change and send, which actions require approval, and who can review its activity or stop it.

An approval button helps only when the person using it can understand the proposed action and its consequences.

AI-assisted coding: working output still needs review

Purpose: Check the consequences of a flaw before relying on generated code.

AI coding tools can help you explore ideas and build prototypes. A demonstration that works does not establish that the application is secure.

In June 2026, the NCSC recommended matching oversight of AI-assisted development to its risks. Code handling sign-in, sensitive information or credentials needs particular care. Read the NCSC’s guidance on reviewing AI-assisted code.

For prototypes, use dummy information and keep real credentials out of prompts and source code. Before launching something that handles real data, arrange a review by someone who understands the relevant security risks.

If you are learning, ask the tool to explain its choices, check those explanations against documentation and seek support where needed. Reviewing only the visible result leaves important behaviour unchecked.

Try this fictional planning exercise

Your team is considering an AI meeting assistant. Its first task is to turn an uploaded transcript into draft action points. It also requests access to every shared folder and permission to send emails.

Before agreeing, discuss:

• Need: which permissions are necessary for the first task?
• Scope: can the team start with one approved transcript?
• Review: who checks the draft before it is shared?
• Recovery: who can revoke access, and where is the activity recorded?

A reasonable first proposal is to work with a selected transcript, handle it according to the organisation’s data rules, and keep sending under human control. Record any access you cannot justify as a question to resolve before adoption.

1. Check what is connected. Identify each app and why it has access. Investigate anything you do not recognise.
2. Remove access you no longer need. On work accounts, follow your organisation’s process and check dependencies before changing access.
3. Review sign-in and recovery. Look for passkey support and confirm you understand how to regain access if your device is lost.
4. Take one question to your team. Ask: “What access does this tool need, and how will we know if it does something unexpected?”

The useful outcome is a clearer picture of who or what can access your account, plus one concrete improvement or question to follow up.