Avast AI Renewal Scam Looks Polished, but Its Code Gives It Away

Avast became the face of an AI-assisted scam after researchers found a polished, five-device renewal page that concealed an unfinished fraud operation. The Avast AI renewal scam looked professional enough to reassure visitors, yet comments buried in its code exposed signs of automated construction.

Malwarebytes found the French-language page while hunting scams targeting people in Belgium. It claimed an Avast Premium Security subscription had renewed and presented an “Active” badge, payment details, dates, and a cancellation form. The page had no connection to Avast, and no renewal charge existed.

The discovery matters because scammers once revealed themselves through broken layouts, awkward sentences, and obvious branding errors. Generative AI can remove those warnings without improving the underlying operation. In this case, the attractive page was reportedly incomplete, but the same production method can generate more convincing variants for other languages and brands.

The Fake Renewal Page Collected Leads, Not Payments

The page was designed to turn a fabricated charge into a verified contact record for a later phone scam.

The fake renewal notice did not immediately request a password, card number, or bank login. Instead, its cancellation form asked for a full name, an email address, and a Belgian mobile number. That smaller request reduced the amount of friction facing a worried visitor.

A familiar company name provided the initial credibility. The apparent renewal supplied urgency, while the cancellation form offered an easy route out. Together, those elements pushed the victim toward disclosure without presenting an obviously dangerous request.

Malwarebytes researcher Stefan Dasic described the operation after his team encountered the page during scam-hunting activity. The company’s renewal page analysis says the site imitated Avast without permission and targeted a Belgian audience.

The page was written in French and presented the supposed subscription as covering five devices. It also showed a future renewal date and a structured transaction summary. Those details gave the fiction the shape of a normal account-management page.

However, the scammers did not need access to an Avast customer database. Sending the same lure broadly would eventually reach some people who recognized the brand or used its products. Others might respond simply because an unexpected charge felt urgent.

The cancellation form served as a lead-generation mechanism. A name connected to a working email address and mobile number gives a scammer several useful signals. It identifies a reachable person, confirms that the renewal story produced concern, and creates a pretext for direct contact.

That sequence is important. Many people judge a form by the sensitivity of each individual field. A request for a phone number can feel safer than a request for a bank password, even when the phone number unlocks the next stage.

After receiving the form, an operator can call while the fake renewal remains fresh in the victim’s mind. The caller can claim to represent billing or technical support and offer to reverse the nonexistent payment. The earlier form submission makes the call appear expected.

This is a form of social engineering, meaning manipulation that persuades a person to take an unsafe action. The page does not need to defeat encryption or exploit a software vulnerability. It needs to establish just enough trust for the victim to continue.

The reported page had a major operational weakness: its form did not transmit the entered information anywhere. That finding suggests the campaign was unfinished, abandoned, or waiting for another operator to connect its collection system.

Yet the incomplete form does not make the design harmless as a model. A finished copy would need only a working destination for submitted data. The visual interface, localized text, invented billing details, and emotional pressure were already in place.

The original security report therefore captured more than an isolated fake page. It documented a reusable production approach that can separate page creation from the later operation of the fraud.

The Avast AI Renewal Scam Left Clues in Its Own Code

The strongest evidence of AI assistance came from material that ordinary visitors would never see.

Researchers found two notes inside the page’s source code. Written in polite French, they addressed the apparent commissioner and explained that the form still required a real submission process. The notes resembled a coding assistant summarizing unfinished work for its user.

That handoff-like language was only one clue. The code also contained unused styling for a section that had been removed. Such remnants can appear when someone generates a page in stages, changes its structure, and skips a final cleanup.

The visible copy created another inconsistency. It was grammatically correct and extended across several paragraphs, but it remained generic. It described broad security benefits without identifying recognizable Avast functions such as firewall, virtual private network, or ransomware protection.

That combination is characteristic of low-effort AI-generated phishing. The output can be fluent, visually orderly, and superficially complete while missing the product knowledge that a real account page would contain.

Two pieces of text would also have appeared as gibberish in a browser, according to the research. That suggests the builder had not completed even a basic visual check. The page looked polished at a glance, but its internal quality control was weak.

These artifacts do not prove which tool produced the code. Generative systems do not normally add a reliable watermark to HTML, CSS, or promotional copy. Humans can also leave unused styles, generic prose, and instructional comments behind.

Malwarebytes therefore framed AI involvement as a supported assessment, not a forensic certainty. That distinction matters because defenders can easily overstate attribution when familiar stylistic patterns appear in malicious code.

An AI-like comment cannot identify a specific model, account, or operator. It also cannot establish whether a model wrote the entire page. A scammer might generate the layout, rewrite selected text, or ask an assistant to repair only one component.

The phrase “AI-generated” can hide those different workflows. A page may contain generated copy, generated code, copied brand assets, and manual changes in the same file. Treating the result as entirely automated would simplify a messier production process.

Still, the evidence supports a narrower conclusion. An operator with limited web-development ability can now obtain a credible first draft through ordinary conversational instructions. The unfinished notes suggest that the system may even have told the operator exactly what remained to be connected.

That lowers the skill barrier without removing the need for criminal planning. The scammer still needs distribution, hosting, a data destination, phone operators, and a way to extract money. AI makes one segment easier and faster.

The abandoned material also exposes a recurring weakness in generated work. A model can produce a convincing interface while leaving dead code, vague wording, or untested behavior. Those faults create signals for investigators, even if victims never see them.

Security teams should preserve complete page source when analyzing suspicious sites. Screenshots capture the victim-facing design, but HTML comments and unused components can reveal the development process. They may also connect apparently separate templates through repeated language or structure.

The Avast AI renewal scam is notable precisely because its surface and internals told different stories. The surface projected competence. The code showed an unfinished handoff, missing functionality, and evidence that no one had completed a careful review.

Polished AI Phishing Breaks the Old Visual Test

A professional appearance no longer offers meaningful evidence that a billing or support page is legitimate.

For years, consumer guidance emphasized spelling mistakes, distorted logos, strange grammar, and crude page design. Those signs remain useful when they appear, but their absence cannot establish trust.

Generative AI can write fluent French, Dutch, German, or English on demand. It can organize transaction details, produce responsive layouts, imitate account dashboards, and rewrite a lure for another brand. None of those outputs require privileged access to the impersonated company.

This changes the economics of localization. A criminal no longer needs a fluent writer and a front-end developer for every target market. One template can be adapted through prompts, then paired with copied images and a new domain.

The result is not necessarily sophisticated code. It is sufficient code delivered quickly. Fraud operations benefit when they can test many brands, languages, and messages while discarding weak variants at little additional effort.

That volume can pressure email providers, hosting companies, domain registrars, browser-security services, and the brands being impersonated. Each organization sees only one part of a campaign that can change its presentation faster than traditional blocklists update.

Google has described related changes across the wider scam market. Its June 2026 online scams advisory noted that attackers use familiar brands and trusted cloud services to obscure phishing infrastructure. The company also reported fake renewal notices delivered through calendar invitations.

The techniques vary, but the defensive problem is similar. People often rely on surface cues when they lack time to investigate a message. Attackers can manufacture those cues faster than they can establish legitimate relationships.

AI-generated phishing also makes inconsistency less dependable as a warning. A message and landing page can now share the same tone, terminology, colors, and invented transaction details. The entire journey may feel coordinated despite being fraudulent.

The better test examines context rather than presentation. Did the person expect this communication? Does the claimed transaction appear in the real bank or card account? Can the subscription be confirmed through an independently opened app?

A legitimate-looking cancellation page cannot answer those questions by itself. It remains part of the claim being tested. Using a button, link, or phone number supplied by the message keeps the victim inside the attacker’s constructed environment.

The supposed renewal should instead be checked through a known account route. The user can open the vendor’s official application or type a verified address independently. If no transaction exists, there is no urgent payment to reverse.

This approach works even when a fake page contains perfect grammar. It also survives future improvements in generated images, personalized wording, and interactive interfaces. Context is harder to fabricate across independent systems than a polished webpage is.

Organizations must adjust training accordingly. Showing employees a gallery of badly written phishing emails can create the wrong lesson. The practical skill is recognizing an unexpected workflow and verifying it through a separate channel.

That shift does not mean visual inspection has become useless. Incorrect domains, broken interface elements, and generic feature descriptions still matter. They should trigger scrutiny, but a clean appearance should never end it.

The Avast AI renewal scam demonstrates the reversal clearly. The most suspicious material was not an ugly typo on the screen. It was polite, competent-looking code commentary hidden beneath a page that appeared ready for customers.

The Dangerous Part Begins After the Form

The renewal page is best understood as the first stage of a human-operated support fraud, not the complete attack.

Once a victim provides contact details, a scammer can call with information that makes the conversation feel connected to the earlier request. Knowing the person’s name, email address, and stated wish to cancel creates an immediate advantage.

The caller can claim that a refund requires remote access. Remote-access software is legitimate technology that allows one computer to control another, but scammers misuse it to view files, manipulate displays, and interfere with financial accounts.

The fictional renewal creates a natural explanation for every next step. The operator may say that account verification is required, that the cancellation failed, or that an incorrect refund appeared. Each claim extends the same invented billing problem.

The US Federal Trade Commission describes this pattern in its support scam guidance. Renewal notices can lead victims to spoofed sites and requests for remote access, followed by pressure involving bank details or difficult-to-reverse payments.

The FBI likewise warns that criminals impersonate technical or customer-support staff, including representatives handling virus-software renewals. Its technical support guidance advises people never to grant device access to an unsolicited support caller.

This is why a form requesting only contact details should not be dismissed as minor data collection. The submitted fields help bridge an anonymous web lure into a live conversation, where an operator can adapt to hesitation in real time.

A caller can answer questions, repeat familiar account details, and increase pressure when the victim tries to leave. That interactive control is harder to achieve through a static page alone.

The call also moves the victim away from the cautious behavior used for suspicious email. People who ignore an unknown link may still trust someone who knows the reason they requested a cancellation minutes earlier.

Historical cases show how costly that progression can become. In one documented Norton-themed scheme, a victim received a fake renewal message, called the supplied number, and granted access to a computer. The operator then manipulated the appearance of the victim’s bank account.

Court records described in a renewal fraud case showed how the invented refund became a reason to transfer money. The initial subscription story was not the final theft method. It was the door into a more elaborate financial deception.

The scale of support fraud makes that distinction urgent. The FBI’s 2025 cybercrime report recorded 47,794 tech-support complaints and more than $2.1 billion in reported losses. Those figures cover a broad category, not this specific page or campaign.

They nevertheless show that the surrounding playbook is established and financially consequential. AI did not invent renewal fraud, remote-access abuse, or refund manipulation. It can make the opening stage cheaper to produce and harder to reject visually.

The unfinished Belgian page therefore represents a capability test as much as an active threat. Even if that particular form never captured data, a completed version could feed an existing call-center workflow.

Defenders should monitor the whole chain. Blocking a domain helps, but so does detecting remote-access software launched after an unusual message. Banks can watch for suspicious transfers, while support teams can warn customers about active impersonation campaigns.

People who already submitted a form should expect a follow-up call. They should avoid engaging, block the number, and treat any claimed refund process as hostile. If remote-access software was installed, they should disconnect the device and seek trusted assistance.

Passwords should be changed from a separate, trusted device when account access may have been exposed. Financial institutions should be contacted directly through verified numbers if payment details or online banking appeared during the interaction.

The crucial lesson is that the visible page may ask for very little because the operator plans to request much more later. Low-friction data collection is not the end of the fraud. It is the beginning of a better-informed conversation.

AI Lowers Production Costs but Does Not Guarantee Competence

The campaign shows a tradeoff between faster scam creation and the new forensic mistakes created by unreviewed generated output.

A person with limited coding experience can ask an AI assistant for a branded subscription page, localized text, a responsive form, and visual status indicators. That collapses several production tasks into one conversational workflow.

The same operator can request new versions for different countries or antivirus brands. The underlying emotional mechanism remains unchanged: present an unexpected charge, create urgency, and offer a cancellation path that benefits the scammer.

This scalability is the strongest security concern. A single polished page is manageable. A stream of short-lived variations can overwhelm manual review and make brand-specific warnings expire quickly.

However, the Belgian page also displayed the limits of generation without testing. Its form was disconnected, obsolete styling remained in the code, and visible text reportedly contained rendering errors. The builder had presentation but not operational readiness.

That tension should shape how analysts discuss AI-enabled cybercrime. Models can lower a barrier without making every attacker highly capable. Automation often creates more attempts, including poorly finished ones, rather than uniformly advanced attacks.

The distinction affects defensive planning. Organizations should not assume every generated page uses a novel exploit or advanced evasion. Many will remain ordinary social-engineering sites wrapped in better design.

At the same time, defenders should not rely on incompetence continuing. Connecting a form to a collection endpoint is a routine step. Future versions can remove comments, test localization, and replace vague copy with real product terminology.

Attackers also learn from takedowns. If researchers identify courteous handoff comments as a fingerprint, template sellers can instruct models to remove them. Static detection rules based on one phrase will age quickly.

Broader patterns offer stronger signals. Those include newly registered domains, mismatches between the claimed brand and hosting infrastructure, cloned page assets, unusual form destinations, and contact collection without a genuine account session.

Behavioral analysis can also focus on the campaign sequence. A message announces a charge, a page requests a phone number, and a caller promotes remote access. Each stage supplies evidence even when the visual content changes.

AI may help defenders analyze those relationships as well. Security systems can cluster similar layouts, compare code structures, and identify repeated infrastructure. Yet automated classification needs human review because AI attribution remains uncertain.

The skeptical question is not whether generative AI influenced this page. The more important question is whether it materially increased the reach, conversion rate, or operational scale of the campaign. The available reporting does not provide those measurements.

No public evidence establishes how many people saw the page, submitted information, received calls, or lost money through this specific operation. The page’s nonfunctional form suggests its direct impact may have been limited when researchers found it.

That uncertainty should prevent inflated claims about a new generation of autonomous fraud. There is no indication that an AI system independently selected victims, operated calls, moved money, or managed the full campaign.

What the evidence does show is narrower and still significant. AI assistance appears to have helped produce a credible localized interface without the level of coding skill previously associated with that result.

The danger comes from combining that capability with an existing criminal supply chain. Template developers, traffic brokers, domain operators, call agents, and money handlers do not need to belong to one organization. A reusable page can pass between them.

The unfinished notes may even support the possibility that the page was a template awaiting sale or deployment. Malwarebytes presented that as an interpretation rather than a confirmed business model.

This uncertainty is a useful reminder for defenders. The first discovered sample may be a draft, not the final campaign. Blocking it is necessary, but tracking its design patterns across other domains and languages can reveal what follows.

What Defenders and Consumers Should Watch Next

The next evidence should reveal whether this was an isolated draft or an early sample of a repeatable multilingual scam kit.

The first signal is template reuse. Researchers should watch for the same layout, code structure, generic benefit language, or form design appearing under other antivirus brands. Versions in Dutch or German would support the claim that rapid localization is part of the model.

Reuse would strengthen the assessment that AI reduced production costs across a campaign. If no related pages appear, the Belgian sample may remain a one-off experiment or abandoned commission.

The second signal is a completed collection and call workflow. Future pages may connect their forms to working endpoints, remove explanatory comments, and initiate fast follow-up calls. That progression would show that operators corrected the weaknesses documented in the first sample.

Users can respond by treating unexpected renewal notices as claims requiring independent verification. Check the relevant bank or card account directly. Then open the security vendor’s official app or website without using the message’s link.

A missing transaction is decisive. There is no payment to reverse if the claimed charge never appeared. The safest response is to report and delete the message, not begin a cancellation process inside its page.

The third signal is a change in defensive guidance and detection. Security vendors, email providers, browsers, and banks should move beyond grammar-based warnings toward context and behavior. Training should explain why a polished page can still be fraudulent.

Organizations can also review whether employees may install remote-access tools without approval. Alerts should connect installation events with recent phishing reports, unusual browser visits, or calls to help desks about unexpected renewals.

Consumers should remain cautious when a form mainly wants a phone number. A company managing a real subscription already has an authenticated account relationship. It should not need an unauthenticated landing page to discover who is canceling.

Anyone receiving a follow-up call should end it and contact the company through a verified channel. Caller ID, knowledge of a submitted form, and a professional tone do not establish legitimacy.

The Avast AI renewal scam does not prove that every polished phishing page was generated by a model. It does prove that visual quality has become a weaker trust signal, while workflow verification has become more important.

The decisive question is no longer whether a page looks like professional work. Ask whether the claimed event exists outside that page, whether the contact path came from an independent source, and whether the requested action fits a legitimate account process.

That habit protects against both AI-generated phishing and older scams built by hand. It also remains useful as templates improve and their internal fingerprints become harder to find.

If an unexpected renewal appears, pause before trying to cancel it. Verify the transaction independently, avoid supplied phone numbers, and never grant remote access to an unsolicited caller. The safest way out of a fake charge is to confirm that it never existed.