AI-powered phishing is changing the cybersecurity landscape

The larger half is where these campaigns now point. Convincing communication is increasingly paired with credential theft, collaboration platforms, QR codes, remote-access tools and abuse of the authentication process itself. The practical question has moved on. It is no longer can our people spot a suspicious email, but what happens when an attacker convincingly impersonates someone our people already trust?

One clarification before the evidence, because the two often get conflated: AI did not invent these techniques. Help-desk impersonation, QR-code lures and remote-access abuse all pre-date the current generation of language models. What AI changes is the cost, speed and plausibility of the human-facing part. The techniques below are growing alongside AI-assisted phishing rather than because of it, and it is worth keeping the two apart when deciding what to do about them.

What the current reporting shows

Those three figures come from two separate bodies of incident data, and they point the same way. Attackers are getting in through people, and once in, they are increasingly working with valid credentials rather than malware.

The phishing attack surface has expanded

The traditional model was linear and easy to teach: an email arrives, a link is clicked, a password is entered. Defences were built around that shape — filter the mail, inspect the link, train people to check the address bar. Modern campaigns arrive through more channels, and they converge on something more useful than a single password.

Every route on the left ends in the same place. That is the important shift: the objective is no longer a password on its own but a working identity — and increasingly a live session, which sidesteps the login step altogether.

When IT support becomes the attack

The clearest published example of this comes from Microsoft. In reporting published in May 2024, Microsoft Threat Intelligence described the financially motivated group it tracks as Storm-1811 misusing Quick Assist, Windows' built-in remote-assistance tool, in social-engineering attacks that ended in Black Basta ransomware.

The sequence Microsoft documented is worth understanding because so little of it is technical. Targets were first flooded with subscription sign-up mail — an email-bombing technique Microsoft calls link listing. The attackers then made contact offering to help with the very problem they had created, by voice call and, from May 2024, through Microsoft Teams using tenants named to look like a help desk. Having established the role, they talked the user through granting a Quick Assist session, then used that access to pull down further tooling, steal credentials through an adversary-in-the-middle kit, establish persistence and eventually deploy ransomware.

Microsoft is explicit that Quick Assist was not compromised. It was used exactly as designed, by a user who believed they were being helped. The vulnerability being exploited was the working assumption that somebody presenting as IT support is IT support — and that assumption is one most organisations have actively trained into their staff.

The practical control is a verification habit rather than a technology. Staff need one internal channel they know how to reach — a published extension, a service-desk number in the intranet, a named manager — and permission to use it without feeling awkward. Unexpected contact should be verified through that channel before anyone:

• Grants a remote-control or screen-sharing session.
• Installs software or a browser extension.
• Runs a command, script or troubleshooting step they were sent.
• Approves a multi-factor prompt they did not personally trigger.
• Discloses a password, one-time code or recovery code.
• Changes a security, MFA or account-recovery setting.

The cultural half matters as much as the procedural half. If verifying is treated as obstructive, people will stop doing it under pressure, which is precisely when it counts.

ClickFix: when the user becomes the installer

ClickFix inverts the usual malware-delivery problem. Rather than smuggling a payload past a mail gateway and an endpoint agent, the attacker presents a plausible obstacle — a failed CAPTCHA, a document that will not render, a meeting that will not join — together with instructions to fix it. The instructions have the user copy a command and run it themselves.

1. A convincing problemA fake verification step, error message or broken document.
2. Helpful instructionsA short fix, usually to paste something into Run, PowerShell or Terminal.
3. The user executes itThe command runs with the user's own rights, from a trusted session.
4. CompromiseA payload is retrieved and the attacker has a foothold.

It works because the victim performs the step that security controls are usually watching for, and performs it deliberately. ESET's threat research for the period December 2025 to May 2026 recorded a 108% increase in ClickFix detections against the preceding six months — the technique roughly doubled in half a year.

That is worth stating carefully rather than absolutely. Support teams do occasionally supply commands, and some environments run genuine scripted fixes. The distinction is provenance: a command that arrives with an unexpected contact, or from a web page presenting an error, is not the same as one issued through a channel the user already knows how to verify. If your own support processes require users to run commands, that is worth revisiting — it trains exactly the behaviour this technique depends on.

Quishing: phishing beyond the link

QR-code phishing has reached record levels in ESET's telemetry, and around 11% of the phishing emails it detected in the period carried a QR code. The appeal to an attacker is structural rather than clever. A QR code hides its destination until it is scanned, so the habit of hovering over a link does not apply. More importantly it usually moves the victim from a managed corporate device onto a personal phone — off the endpoint agent, off the corporate network path, and away from the browser controls the organisation actually operates.

The lures imitate exactly what people expect to receive: a Microsoft 365 sign-in, a shared document, a password-expiry notice, an MFA enrolment step, an invoice or a delivery notification.

There is a South African detail here that cuts against complacency. ESET reports that quishing has entered the top ten threats for South Africa and for Africa — but that local detection, at around 2.1%, sits well below markets such as North America at roughly 12.4%. The natural reading of a low number is that the problem is smaller here. The more useful reading is that the technique has room left to grow locally, and that the window to prepare is now rather than after the numbers converge.

Identity is becoming the new perimeter

None of this makes the network boundary irrelevant. Firewalls, segmentation and endpoint controls all still do work that nothing else does. But the centre of gravity has moved. Most organisations now run across Microsoft 365, a spread of SaaS applications, cloud infrastructure, mobile devices, remote workers and third-party integrations. In that shape, a valid identity is not a step towards access — very often it is the access, and it arrives looking entirely legitimate.

Cisco Talos' incident-response trends for Q2 2026, published in July 2026, put numbers to this. Phishing was the primary means of initial access in over half of their engagements, up from about a third the previous quarter. Authentication abuse appeared in 65% of engagements, against 35% the quarter before — with attackers defeating multi-factor authentication using adversary-in-the-middle proxies, stolen session tokens, MFA fatigue and self-enrolled devices.

Read those together and the pattern is not subtle. The way in is a person; the way onward is their identity.

The identity attack chain, and where it breaks

Set out as a sequence, the path is consistent enough to plan against — and each stage has a control that interrupts it. No single one of them is decisive, which is the point: the chain breaks wherever it is thinnest for the attacker, not wherever an organisation spent the most.

1. Social engineeringConvincing contact by email, Teams, phone or QR code.
   ControlSecurity awareness & email security
2. Credential or session compromiseA password is captured, or a live session token is stolen and replayed.
   ControlPhishing-resistant MFA
3. A valid identityThe attacker now authenticates as a real user, from their own device.
   ControlConditional access & device compliance
4. Cloud and Microsoft 365 accessMail, files, chat and connected applications become reachable.
   ControlIdentity monitoring & risky sign-in detection
5. Privilege and lateral movementRights are escalated or reused to reach systems the first account could not.
   ControlLeast privilege & admin separation
6. Business data and systemsData is taken, altered or encrypted; finance and operations are affected.
   ControlEndpoint protection, logging & response

Passwords and basic MFA are no longer the end goal

To be unambiguous, because this argument is often mangled: multi-factor authentication is enormously worthwhile. Any reasonable second factor is a substantial improvement on a password alone, and an organisation without MFA should be enabling it rather than reading ahead. The point is that not all factors resist the same attacks.

1. Password onlyA shared secret. Reusable, phishable, and traded in bulk.
2. App or SMS MFAA large improvement, but a code can be relayed and a prompt can be approved under pressure.
3. Phishing-resistant MFACryptographically bound to the real site, so a proxy in the middle has nothing to relay.
4. Passkeys and FIDO2No shared secret exists to steal, and nothing useful is typed.

Codes and push approvals share a weakness: they can be forwarded by a person. An adversary-in-the-middle page can relay a one-time code in real time, and a user worn down by repeated prompts may approve one to make it stop. FIDO2 credentials, passkeys and Windows Hello for Business remove that possibility, because the credential is bound to the legitimate site and never leaves the device in a form anyone can pass on. Certificate-based authentication serves a similar purpose in environments already running a certificate authority.

This is also becoming a scheduling matter rather than a strategy debate. Microsoft's own documentation states that from 1 September 2026, passkeys become the default authentication experience in Microsoft Entra ID. Three qualifications matter for planning, and are frequently reported inaccurately:

• The automatic enablement targets users who are enabled for SMS or voice in the authentication methods policy — not every user in every tenant.
• Those users are auto-enabled for passkeys and nudged to register at their next MFA sign-in. By Microsoft's documentation the prompt can be snoozed without limit, so September is a push rather than a hard cut-over.
• A temporary opt-out does exist for the September window, set through the authentication methods policy. It expires on 1 February 2027, when Microsoft-provided SMS and voice are retired — and from that date a user whose only method is SMS or voice faces a blocking passkey registration prompt, with no opt-out for any tenant.

For most South African organisations the practical takeaway is a date. There is a defined window to move people onto phishing-resistant methods deliberately, or to have it happen to them at a less convenient moment.

Build an identity-first defence

None of the following is exotic. It is the set of controls that keeps appearing on the right side of incident reports:

Before you trust the request

• Who is contacting me, and through which channel?
• Was I expecting this contact?
• Why does this person need this access?
• Can I verify the request through a number or channel I already know?
• Am I being asked to run, install or paste something?
• Am I being hurried, or told this is urgent?
• Does this need my password, a one-time code or an MFA approval?
• Is this genuinely our IT team, or someone presenting as our IT team?

The Define IT perspective

Cybersecurity is no longer only about protecting a network. People, identities, endpoints, applications, networks and data have to be treated as one connected environment, because that is how an attacker treats them — entering through a person and moving through whatever the resulting identity can reach.

Define IT works with organisations across that whole surface: cybersecurity covering identity, endpoint, email and network security along with vulnerability management and monitoring; Microsoft solutions for Entra, Microsoft 365 and the access controls around them; managed IT services for the monitoring, patching and maintenance that keep controls working after the project ends; and backup and disaster recovery for the point at which prevention has already failed.

Secure your people. Protect your identity. Defend your business.

Key takeaways

• AI has not created these techniques — it has made the human-facing part cheaper and more convincing. Treat AI-assisted phishing and identity attacks as related but separate problems.
• The objective has moved from a password to a working identity, and often to a live session that bypasses the login step entirely.
• Give staff one verification channel they know how to reach, and make using it socially acceptable — especially under time pressure.
• Phishing-resistant methods differ from MFA in kind, not degree: a passkey has no shared secret a person can be talked into passing on.
• Microsoft's Entra passkey default begins 1 September 2026 with a snoozable prompt and a temporary opt-out; the hard enforcement date is 1 February 2027.

Related Define IT services

• Cybersecurity
• Microsoft Solutions
• Managed IT Services
• Backup & Disaster Recovery