The barrier to launching a phishing campaign used to be technical skill. That barrier no longer exists.
For most of the history of phishing, launching a convincing attack required knowing how to code. You needed HTML and CSS to clone a login page. You needed server infrastructure to host it. You needed JavaScript to capture and exfiltrate credentials. You needed enough operational knowledge to configure DNS, obtain SSL certificates, and avoid obvious detection signatures.
The skill floor was not impossibly high, but it was high enough that the population of people who could build effective phishing campaigns was limited to those with at least a moderate level of technical competence.
That constraint shaped the threat landscape for two decades. Phishing kits lowered the barrier by packaging infrastructure into reusable templates, and phishing-as-a-service platforms lowered it further by offering turnkey operations for a subscription fee. But both still required the attacker to understand what they were deploying. The tools made it easier. They did not make it effortless.
Vibe coding has made it effortless.
What vibe coding actually enables
The term was coined by Andrej Karpathy, co-founder of OpenAI, in early 2025 to describe a style of software development where the user describes what they want in plain language and an AI model generates the code. The user does not need to understand the code, review it, or even look at it. They describe the outcome. The platform builds it.
Platforms like Lovable, Replit, Bolt.new, and Vercel’s v0 have turned this concept into a massive market. Lovable reached $100 million in annualized revenue within eight months of launch. Replit is valued at roughly $9 billion. These are not fringe tools. They are mainstream development platforms with millions of users, and they can produce fully functional, professionally designed web applications from a few sentences of natural language input.
The security implications became public in April 2025 when Guardio Labs published research they titled “VibeScamming.” The researchers tested whether popular AI platforms could be prompted to build complete phishing campaigns. The results were stark.
Lovable scored 1.8 out of 10 on Guardio’s resistance benchmark, meaning it was the most susceptible to generating malicious output. It did not just produce a convincing Microsoft login clone.
It hosted the page on its own subdomain (*.lovable.app), generated credential storage with a fully functional admin dashboard for reviewing stolen passwords, IP addresses, and timestamps, and redirected victims to the real office.com after harvesting their credentials. The researchers did not ask for the admin dashboard. Lovable generated it automatically, anticipating the user would want to review captured data.
Proofpoint subsequently tracked live campaigns where criminals used Lovable to host credential harvesting pages impersonating Microsoft, UPS, and cryptocurrency platforms.
In February 2026, RavenMail documented a campaign where Lovable was not even hosting the final phishing page. It was being used as a multi-stage redirector, routing victims through a lovable.app URL before sending them to a separate credential harvesting site. The lovable.app domain passed every reputation check. Email gateways and Microsoft Defender did not flag it.
The traditional barriers that prevented non-technical threat actors from building sophisticated phishing campaigns, knowledge of HTML, CSS, JavaScript, server configuration, deployment, have been eliminated. A credential harvesting page that would have taken a moderately skilled developer several hours to build can now be produced in minutes through a conversational interface.
Why this is not the same problem as LOTS
Organizations familiar with Living Off Trusted Sites attacks might see parallels here. Both involve attackers operating on legitimate infrastructure. Both degrade the value of domain reputation as a detection signal. But the vibe coding problem is structurally different in ways that matter for defense.
LOTS attacks involve an attacker who builds phishing infrastructure independently and then hosts it on a trusted platform like Cloudflare, AWS, or GitHub. The attacker brings the skill. The platform provides the trust. The hosting is passive.
Vibe coding attacks involve the platform itself generating the phishing infrastructure. The attacker describes what they want. The AI builds it, deploys it, and in Lovable’s case, even provides the operational tools to manage the campaign.
The platform is not just a host. It is the weapon. And it requires no technical skill to operate.
This distinction matters because it changes who can be an attacker. LOTS expanded the hosting options for technically skilled operators. Vibe coding expands the population of operators to anyone with a laptop and a prompt.
The addressable market for phishing-as-a-service was already large. Vibe coding makes the concept of “as-a-service” redundant, because the general-purpose AI platforms provide the service without being designed for it.
Our detection data reflects the early signal of this shift. Allure’s SPOOF ’26 annual threat report categorizes “vibe coding platforms” as a distinct infrastructure source for the first time, documenting 376 alerts from this category in 2025. Vercel alone accounted for 341 of those alerts, nearly as many as the entire traditional website hosting category combined. That number will grow. The platforms enabling it are valued in the billions, and a category that did not exist in our 2024 data generated meaningful volume in its first year.
What this breaks in the defensive model
Most brand protection and anti-phishing tools detect threats by looking at infrastructure signals: domain names, registration age, hosting reputation, SSL certificate patterns. These signals are effective when attackers operate on infrastructure that looks different from legitimate business operations. A freshly registered domain with a name similar to your brand, hosted on an unfamiliar provider with a free SSL certificate, generates useful signal.
A phishing site built on Lovable and hosted at a random-string.lovable.app address generates none of those signals. The domain is legitimate. It is not new. It has strong reputation. The SSL certificate is valid and issued by the same authority that covers every other Lovable project. The hosting infrastructure is the same infrastructure used by tens of thousands of legitimate applications.
Our research confirms how broadly this pattern has taken hold. Only 7% of domains used in phishing attacks against financial institutions are less than 30 days old. Forty-one percent are over five years old. The conventional advice to watch for newly registered lookalike domains sees less than a tenth of the actual threat surface, and vibe coding platforms are accelerating the move away from attacker-controlled infrastructure entirely.
Domain monitoring is not worthless. It still catches the subset of attacks that use deceptively named domains. But as a primary detection strategy, it is increasingly blind to how attacks are actually built and deployed. When the attacker’s domain is lovable.app or a Vercel deployment URL or a Replit project link, the domain is not the signal. The content is.
The platform breadth tradeoff: external cybersecurity vs. purpose-built brand protection
ZeroFox built its business as a broad external cybersecurity platform. Brand protection sits alongside threat intelligence, attack surface management, executive protection, dark web monitoring, and physical security intelligence. For a CISO looking to consolidate external risk visibility under a single vendor, the breadth is appealing.
The tradeoff is depth. When brand protection is one product line competing for engineering resources, roadmap attention, and go-to-market investment alongside five or six others, buyers should ask what that means for the speed at which the brand protection capability evolves. User reviews have surfaced this in practice: takedown success rates vary, social media coverage has documented gaps on platforms like LinkedIn, Facebook, and Instagram, and multi-week takedown timelines are not uncommon.
Allure Security does one thing: brand protection. The entire engineering team, SOC, and product roadmap serve that mission. When new attack surfaces emerge, whether it is phishing hosted on npm registries, impersonation campaigns running on legitimate cloud infrastructure, or QR code phishing bypassing email security entirely, the response is immediate because nothing else competes for the roadmap. Purpose-built means every detection model, every SOC workflow, and every product investment serves one outcome: finding brand impersonation and stopping it.
The question is not which model is universally better. It is which one matches your organization’s actual need. If you need a single platform for external cyber risk across many dimensions and your team has the capacity to operate it, ZeroFox’s breadth has value. If brand impersonation is the specific problem keeping you up at night and you need it solved without adding headcount, depth wins.
What detection needs to look like now
The shift toward attacker infrastructure that is indistinguishable from legitimate business operations has been building for years. Vibe coding accelerates it by removing the last remaining constraint: technical skill. The detection response has to match.
Content-based analysis, examining what a page says, how it behaves, and what it is designed to collect, works regardless of where the page is hosted. A credential harvesting form impersonating a bank login is identifiable by its content and intent, not by its domain name or hosting provider. Visual brand replication, form behavior, and exfiltration patterns are signals that persist even when every infrastructure indicator looks clean.
This is why Allure Security’s detection architecture scans 1.4 billion web pages daily using computer vision and natural language processing rather than relying on domain-level signals. When a phishing page is built on Lovable, hosted on Vercel, and served through a legitimate CDN with a valid certificate, the only way to identify it as malicious is to look at what is on the page. Infrastructure tells you nothing. Content tells you everything.
The speed dimension matters here too. Vibe coding compresses the time between “attacker has an idea” and “phishing site is live” from hours to minutes. Our research shows that 75% of phishing victims arrive within ten hours of a site going live, with a quarter arriving in the first four hours.
When site creation takes minutes and victim exposure concentrates in hours, detection and blocking that operates on longer timescales concedes the window where protection is still possible. Allure’s approximately 15-minute median detection-to-blocking time operates within that window. Most takedown processes do not.
The Bottom Line
The phishing threat model that most security programs are built around assumes a technically skilled attacker with purpose-built infrastructure. Vibe coding invalidates both assumptions simultaneously. The attacker does not need technical skill. The infrastructure is not purpose-built.
And the platforms generating this new category of threat are growing faster than almost any category of software in history. Organizations whose brand protection relies on watching domain registrations and scoring hosting reputation are defending against yesterday’s attacker with yesterday’s tools. The one that matters today is building phishing sites by describing them in plain English, on platforms that millions of legitimate businesses also use, and the only way to find those sites is to look at what is on the page.
Key Takeaways
Vibe coding is a style of software development where users describe what they want in natural language and an AI platform generates the code. Platforms like Lovable, Replit, and Bolt.new have made this accessible to anyone. For brand protection, the implication is that building a convincing phishing site no longer requires technical skill. Anyone with a prompt can produce a fully functional credential harvesting page in minutes, dramatically expanding the population of potential attackers.
Guardio Labs’ VibeScamming research found that Lovable could be prompted to generate pixel-perfect Microsoft login clones, host them on its own subdomain, and even provide admin dashboards for managing stolen credentials. Proofpoint and RavenMail have documented live campaigns using Lovable as both a hosting platform and a multi-stage redirector. Because the resulting sites live on legitimate domains like lovable.app, they bypass domain reputation checks and email security filters.
Domain monitoring detects threats by identifying suspicious domain names, new registrations, or lookalike URLs. A phishing site hosted on lovable.app or a Vercel deployment URL has a legitimate domain, strong reputation, valid SSL, and established hosting infrastructure. None of the signals domain monitoring relies on are present. Detection requires examining page content, visual brand replication, and credential harvesting behavior rather than infrastructure metadata.
Organizations should ensure their brand protection includes content-based analysis that identifies impersonation by examining what a page does rather than where it is hosted. Detection speed also matters: with site creation taking minutes and 75% of victims arriving within ten hours, blocking needs to happen within the first hour to protect the majority of potential victims.
