Phishing in 2026: Modern Tactics and Defenses
Quishing, adversary-in-the-middle kits, AI-written lures, and ClickFix scams have changed phishing. Here is what works in 2026 and how to defend against it.
By Velloreth
Phishing has not changed its goal — get a human to hand over credentials or approve a session — but the delivery mechanics of 2026 look nothing like the misspelled bank emails of 2016. QR codes, adversary-in-the-middle proxy kits, AI-personalized lures, and fake “support” instructions have taken over. Understanding how each one works is the difference between a defense that holds and one that gets bypassed on the first real attempt.
What the 2026 numbers look like
A few data points worth knowing:
- Microsoft reported detecting and analyzing 8.3 billion phishing threats in Q1 2026 alone.
- QR-code phishing (“quishing”) surged 146% quarter-over-quarter in Q1 2026, and ESET found malicious QR codes in about 11% of all phishing emails in the first half of 2026.
- ESET also tracked a fivefold rise in malicious “AI skills” between March and May 2026 — components that let AI agents access systems, exfiltrate data, and run malware.
- OpenSSL-adjacent research aside, the practical finding from ESET’s H1 2026 report is that attackers are now embedding compromise chains inside AI-generated troubleshooting content (“AI-fix” / ClickFix), with detections of that vector more than doubling between H2 2025 and H1 2026.
None of these numbers mean phishing is unstoppable. They mean the old training material — check the sender, look for typos — covers less of the attack surface than it used to.
Tactic 1: Adversary-in-the-middle kits
This is the most important tactic to understand, because it defeats standard MFA.
A traditional phishing page proxies your credentials to the real site in real time. An adversary-in-the-middle (AiTM) kit goes further: it sits between the victim and the legitimate site, relays the login, captures the resulting session cookie, and hands the attacker a logged-in session. The victim enters a real password and a real TOTP code; both are forwarded correctly. The attacker never needs the password again — they replay the cookie.
What this means in practice:
| Defense | Stops credential phishing | Stops AiTM |
|---|---|---|
| Unique passwords + password manager | Limits reuse damage | No |
| TOTP / authenticator app MFA | Yes (for simple kits) | No |
| SMS one-time codes | Weakly | No |
| Phishing-resistant MFA (passkeys, FIDO2 keys) | Yes | Yes |
| Device-bound session cookies | Partial | Yes, with platform support |
The fix is phishing-resistant authentication. Passkeys and FIDO2 security keys are bound to the origin they were created for, so an AiTM proxy on a lookalike domain cannot get a signature from them. Device-bound cookies (the approach behind Cloudflare’s and Google’s session-bound work, and standardized as the Sec-Session / related proposals in progress) add a second layer: even a stolen cookie is useless on a different device.
If your organization still relies on TOTP as the primary MFA factor, assume your AiTM exposure is open. Prioritize passkeys on high-value accounts — email, code hosts, cloud consoles, identity providers.
Tactic 2: Quishing
Quishing replaces the clickable link with a QR code image, often embedded in a PDF attachment or the email body. The mechanics:
- The QR encodes a URL pointing to a phishing page, frequently a fake Microsoft 365 login, DocuSign request, or MFA prompt.
- Email security tools that scan link text never see the URL — it is pixels.
- Scanning moves the victim from a managed laptop with proxy filtering and endpoint protection to a personal phone with neither.
- Many campaigns chain into AiTM, so even a user who “just scans quickly” can hand over a session token.
Executives are disproportionately targeted. One 2025 analysis found C-level employees received QR-code phishing attempts roughly 40 times more often than the average employee — a plausible-looking DocuSign or IT ticket is exactly the kind of message a busy executive scans without scrutiny.
What works against quishing:
- Decode QR codes at the gateway. Several email security products now extract and scan URLs from QR images inside email bodies and attachments. If your gateway supports it, enable it.
- Train the specific pattern. The message to staff is not “QR codes are bad” — it is “an unexpected QR code in email is an unexpected link. Verify with the sender through a channel you already trust.”
- Keep the phone side covered. Require a managed browser or at minimum a modern mobile OS with Safe Browsing–style protections. The corporate proxy no longer applies on the phone, so the browser is your control point.
Tactic 3: AI-personalized lures
Language models removed the telltale grammar errors. They also made personalization cheap. An attacker can pull a target’s recent conference talk, GitHub activity, or job posting and write a lure that references real details. Research from Eindhoven University in 2025 tested LLM-assisted phishing against real organizations using open-source intelligence and found it competitive with conventional templates.
The practical implication for defenders: “spot the broken English” training is obsolete. What still separates real from fake:
- Context you did not initiate. A message about an account you do not remember creating, or a document you never shared.
- Unusual channel or urgency to authenticate. Nobody legitimate needs your MFA code today, ever.
- Links that do not match the visible domain. Hover or long-press the link, and check the registered domain, not the subdomain soup.
Teach pattern recognition around situation, not grammar.
Tactic 4: ClickFix and “AI-fix”
The newest social engineering variant does not even show a fake login page. The attacker sends a message describing a problem that does not exist — “your subscription failed,” “your meeting recording is corrupt,” “your account has a configuration error” — and instructs the victim to paste a command into their terminal or PowerShell to fix it. The command downloads and runs malware or a loader.
ESET calls the AI-generated variant “AI-fix”: troubleshooting content generated by language models that embeds a compromise chain, with detections more than doubling between H2 2025 and H1 2026.
This works because:
- It exploits the victim’s willingness to help their own system.
- The “instructions” look technical and therefore trustworthy to non-specialists.
- It bypasses email link scanning entirely — there is no malicious URL, just a command to paste.
Defenses:
- Explicit policy: nobody asks you to paste commands from a message. Not IT, not Microsoft, not a vendor. Real support workflows do not use paste-a-command.
- Application controls where feasible. On managed machines, restrict script execution (PowerShell Constrained Language Mode, allow-listed interpreters).
- Report-and-verify path. Give staff a fast way to ask “is this real?” before they act, and make sure the answer comes within minutes, not days.
Tactic 5: MFA fatigue and push bombing
Before a full AiTM kit, attackers sometimes try the cheapest MFA bypass: spam the victim with approval prompts until they tap “Approve” to make it stop. This is push bombing or MFA fatigue, and it still works on organizations that treat any successful MFA as sufficient.
It pairs with a credential-phishing step: the attacker already has the password (from a breach or a simple kit) and is fishing for the second factor. The victim gets repeated notifications at odd hours, approves one out of frustration, and the attacker is in.
Modern platforms have largely mitigated this with number matching (the prompt shows a number the user must type from the login screen) and geographical anomaly warnings. If your MFA provider supports these, enable them. If users can still approve a bare push with no context, update the configuration.
Tactic 6: Business email compromise with a modern costume
Business email compromise (BEC) has always been the highest-damage phishing category, because it targets wire transfers and payroll changes rather than credentials. The 2026 version adds AI drafting, QR codes for “invoice review,” and thread hijacking — replying inside a real conversation after compromising one participant.
BEC defenses are mostly process:
- Dual approval for any payment or bank-detail change.
- Callback verification on a phone number you already have on file, never one in the email.
- Domain-level DMARC enforcement (
p=reject) so attackers cannot send mail from your domain to your own staff.
A pattern worth internalizing: BEC almost never asks you to log in. It asks you to act — pay this invoice, change this payroll account, send this document. That is why credential-focused training leaves people unprepared for it. Drill the payment-verification process specifically, not just login pages.
Testing your defenses without a security team
You do not need an enterprise security awareness platform to find out whether your people would fall for the current tactics. A lightweight internal drill works:
- Pick one tactic. A QR code in a fake invoice email is a good 2026 choice, because it tests both link hygiene and the new quishing surface.
- Send it from a safe sandbox. Use a test domain you control; never simulate against real customers or with real malicious payloads.
- Measure the right thing. Click/scan rate matters, but so does report rate. A team that reports the phish quickly is safer than a team that never clicks but also never alerts you when a real one lands.
- Follow up in the moment. Within a day, share what the email was, why it was fake, and what to check next time. A drill without a debrief teaches helplessness, not vigilance.
- Repeat quarterly with a different tactic. Quishing one quarter, payment-change BEC the next, ClickFix the quarter after. Training that only covers the last drill’s exact template fails on the next variant.
The goal is not a gotcha scoreboard. It is a baseline, a teaching moment, and a report path that people actually use.
The defense stack that holds up
Grouping the above, a working 2026 phishing defense has five layers:
- Phishing-resistant MFA on everything that matters. Passkeys or FIDO2 keys for identity provider, email, code hosts, cloud. TOTP is a fallback at best, not a primary.
- Email authentication done properly. SPF aligned, DKIM signing, DMARC at
p=rejectfor your domains. This stops direct domain spoofing, which is cheap for attackers and expensive for you. - Session protections. Short session lifetimes on sensitive apps, and device-bound cookies where the platform supports them. This blunts AiTM even when a cookie leaks.
- Awareness aimed at current tactics. Not “look for typos” — instead: unexpected QR codes, paste-a-command instructions, requests to approve MFA you did not start, and urgency around payments.
- A fast report path. One address or one button. Every reported phish is a free detection update, and staff who report feel ownership instead of fear.
A note on AI agents
ESET’s finding on malicious AI skills points at the next phase: attackers are not only using AI to write phishing, they are shipping components that let AI agents inside your environment do the exfiltrating. The zero trust framing applies directly — treat agent outputs and tool calls as untrusted input, scope agent credentials to the minimum, and never let an agent hold standing admin rights because “it needs them for the demo.”
Phishing in 2026 is a craft of moving trust. The QR code moves you to a phone. The AiTM kit moves the session. The ClickFix message moves a command into your shell. Every defense that works is one that removes a link in that chain — origin-bound credentials, gateway decoding, script controls, callback verification. Pick the layers that close the chains your people actually face, and revisit them twice a year. The tactics will keep changing; the structure of the defense does not have to.