5 min read

The Right Tool at the Right Time: Inside an Automated Takedown at Netcraft

Facebook logo
Facebook logo
X (formerly Twitter) logo
X (formerly Twitter) logo
LinkedIn logo
LinkedIn logo
Reddit logo
Reddit logo

Overview

At Netcraft, an automated takedown starts long before a takedown request reaches a registrar or hosting provider.

First, the system has to find the threat and retrieve the version of the page the intended victim would actually see. Then it has to classify what is there, distinguish a real attack from legitimate or ambiguous content, capture the evidence, and send a report that the organization responsible for removal can act on. Each of those steps affects whether a phishing site comes down (and how quickly), and calls for different tools.

Netcraft uses more than 600,000 specialist classifiers alongside machine learning, newer AI models, analyst judgment, a global proxy network, and automated workflows. That means a new phishing kit can get its own classifier within hours, boundary cases go to analysts, and evidence is captured from the geography and vantage point targeted by the attacker, so registrars and hosting providers don’t have to recreate the attack themselves. 

Let’s explore how those pieces fit together inside an automated takedown at Netcraft.

One model or 600K

Netcraft’s detection system includes more than 600,000 classifiers designed to recognize individual indicators, such as a brand’s visual identity, a phishing kit’s code, a hosting pattern, or a lure template.

When a new kit appears, a classifier can be added within hours without retraining the system as a whole. Every classifier learns from examples labeled by analysts who have been studying and labeling cybercrime for well over a decade.

But specialist classifiers are only one tool in the system. Netcraft runs vision models to detect logo misuse, speech recognition models to classify phone calls, and frontier models for novel edge case attacks where deterministic classifiers are not enough.

The technology itself becomes meaningful when you look at the job each tool is doing.

What needs to happen

How Netcraft approaches it

Recognize known malicious indicators, threat actor techniques, and brand-specific tactics

Classification trained to recognize malicious patterns and threat actor techniques

Respond to newly observed phishing techniques

Recursive learning from each new tactic, technique, and procedure

Interpret novel edge cases

Models selected for the task, informed by attack patterns observed across brands and refined through analyst expertise

Retrieve attacks hidden from commercially available scanners

Behave like a victim, matching geolocation, device type, and referring ad

Support removal

Package evidence and details into a complete and accurate takedown request

Accuracy depends on the scale and quality of the data 

A reliable classification engine depends on two capabilities: finding and retrieving malicious content and classifying what it finds. Both get better as the scale and quality of the data behind them improves. 

Phishing data ages quickly. Threat actors continually change their evasion tactics, introducing new cloaking methods, phishing kit code, hosting setups, phishing infrastructure, and scams. A public dataset captures what phishing looked like when someone compiled it, and attackers will keep moving at a rapid pace.

Key takeaways

  • Automated takedown is a workflow. Detection, classification, evidence gathering, review, and escalation all affect whether a threat gets removed.

  • Accuracy depends on context as much as classification. Seeing the same content as the intended victim and knowing when to involve an analyst both shape the quality of the final takedown request.

  • Intelligent automation comes from combining different tools across that workflow. Netcraft uses specialist classifiers, broader AI and machine learning, analyst expertise, and automated processes according to the job each one needs to do.

Key takeaways

  • Automated takedown is a workflow. Detection, classification, evidence gathering, review, and escalation all affect whether a threat gets removed.

  • Accuracy depends on context as much as classification. Seeing the same content as the intended victim and knowing when to involve an analyst both shape the quality of the final takedown request.

  • Intelligent automation comes from combining different tools across that workflow. Netcraft uses specialist classifiers, broader AI and machine learning, analyst expertise, and automated processes according to the job each one needs to do.

Netcraft performs roughly one-third of the world’s phishing takedowns, and that volume feeds a cybercrime dataset its analysts have labeled and added to for well over a decade, updated as attacker techniques change.

That analyst-informed threat intelligence helps keep false positives low in Netcraft’s Phishing and Scam Protection Platform, reducing the chance that a legitimate site gets classified as a phishing or brand impersonation attempt, and giving infrastructure providers reason to treat its reports as actionable.

Why classifiers have to see what the victim sees

A classification engine can only assess the content it is given, so what a scanner retrieves sets the ceiling on what classification can achieve.

Before a classifier sees anything, a scanner has to visit the URL and fetch the page. Attackers set up malicious websites to check who’s visiting, looking at things like IP address, device type, browser, and whether the visitor came from a link in a phishing email.

If the visitor looks like a security scanner, the site shows a harmless page, a 404 error, or redirects to the brand’s official site. If it looks like a genuine target, the site shows the lookalike domain and fake login page. A perfect classifier looking at the harmless version will classify what it sees correctly, but the phishing page stays live.

Netcraft’s classifiers analyze the page a victim would see, not the decoy. Its global proxy network helps bypass cloaking and geofencing techniques by retrieving content from the geography and conditions the attacker is targeting.

That means when Netcraft sends a takedown request, the verdict is based on the phishing content shown to the intended victim rather than the harmless version served to a known security scanner.

Analyst expertise strengthens classification

Most classifier verdicts are clear cut, but some phishing pages and suspicious digital assets sit in a gray area where the right answer depends on context the classifier cannot see or techniques no one has seen before

For example, a threat actor might deploy a technique no classifier has been built for yet, obfuscate a page in a way the system has not encountered, or alter a brand’s logo just enough that a vision model no longer recognizes it.

In those cases, when the deterministic classifiers cannot reach a verdict, Netcraft applies probabilistic models, and routes the case to an analyst when those models cannot settle it either.

If a phishing takedown service submits repeated requests based on incomplete or incorrect information, including false positives, a domain registrar or hosting provider may investigate and find nothing malicious. Over time, that can weaken trust in the sender’s reports and affect how future takedown requests are handled.

A false negative has the opposite effect: a real phishing site stays live, potentially appearing in search engines and giving victims more time to hand over credentials, download malware, or expose sensitive information and intellectual property.

That is why Netcraft’s analysts check boundary cases before takedown requests go out, then turn what they find into detection queries and classification rules that apply across every brand the platform protects.

Turning verdicts into evidence a registrar can act on

A domain takedown service cannot remove a phishing page itself. It relies on a domain registrar, hosting provider, platform, or another third party to carry out the content removal. Those third parties cannot act on a system’s verdict alone, so vendors must produce a detailed evidence package that allows them to verify the abuse and follow their own takedown procedures.

To stand up to scrutiny, that evidence also needs to reflect what the attacker is showing its target.

A phishing page cloaked to appear only to mobile users in the UK will show a registrar in the US a harmless page, and the registrar will find nothing to remove. Netcraft captures the page through our global proxy network, so the report documents what a victim sees. 

Registrars do not need to recreate attack conditions to confirm a threat, because the evidence package documents it, including screenshots, timestamps, and obfuscation data. What each report contains depends on what the infrastructure provider needs in order to act.

Every part of Netcraft’s approach contributes to the quality of its reports. Extensive threat data and task-specific models help identify attacks. Our global proxy network helps retrieve the malicious content shown to intended victims. Analyst expertise helps interpret complex cases and improve detection. Together, these capabilities produce clear, well-supported takedown requests.

That combination of proprietary threat intelligence, classification, targeted retrieval, evidence gathering, and automated escalation operates at a significant scale. Consistently supplying that level of detail is what gives registrars and hosting providers reason to treat Netcraft’s reports as worth acting on.

How the automated takedown system performs in practice

Phishing kits, cloaking techniques, and attacker infrastructure will keep changing, as will the models available to defenders. Netcraft’s approach is built to keep incorporating new tools and new intelligence without depending on any one of them.

Today, that system removes phishing URLs in as little as 2.1 hours. Maintaining that performance as attacks evolve means continuing to improve each part of the process: adding classifiers for emerging patterns, applying newer AI models where they help, feeding analyst insight back into detection, and strengthening the evidence that supports removal.

Intelligent automation gives Netcraft room to keep moving as tools improve, without losing focus on faster, accurate takedowns that hold up.

Want more insights like this?

New findings on phishing infrastructure, scam campaigns, and takedown trends - right to your inbox.

Want more insights like this?

New findings on phishing infrastructure, scam campaigns, and takedown trends - right to your inbox.

Want more insights like this?

New findings on phishing infrastructure, scam campaigns, and takedown trends - right to your inbox.

See how Netcraft puts intelligent automation to work across threat detection and takedown.

Get a demo

How do digital risk protection platforms classify phishing sites automatically?

Digital risk protection platforms and domain takedown services classify retrieved pages automatically, though the approach varies by provider. A classifier scores a page against patterns learned from labeled examples, such as a brand’s visual identity, a phishing kit’s code, or a hosting setup. Some platforms rely on one general model, but others, like Netcraft, run many narrow classifiers that can be added or retrained as new threats appear, alongside models selected for specific tasks.

What is a false positive rate in brand protection and why does it matter?

A false positive happens when legitimate sites are incorrectly classified as malicious domains. It matters because takedown requests ultimately go to registrars, hosting providers, app stores, social platforms, or other organizations responsible for removal. Frequent false positives can weaken trust in future reports. Accurate classification helps make takedown requests more credible and keeps legitimate content out of the removal process.

Can a phishing site be taken down without human review?

Sometimes, yes. Netcraft partners with more than 500 infrastructure and service providers to improve takedown times and take down threats before they can go live, which means parts of the process run without a person in the loop on either side. The registrar, hosting provider, or other organization responsible for the infrastructure still decides whether to remove the content. 

Human judgment also enters earlier in Netcraft’s workflow, where most threat classification runs automatically and analysts review boundary cases that depend on context a classifier cannot see.

How does automated takedown fit into a broader digital risk protection program?

Automated takedown is the backbone of a digital risk protection program. Threat actors launch new phishing sites quickly and replace infrastructure after it is removed, so keeping pace means detecting and disrupting at a scale no team can manage request by request. A broader digital risk protection program can monitor threats across domains, websites, social media, mobile apps, app stores, and other external channels, then use validated threat intelligence to prioritize what needs disruption or removal.

For security teams, that intelligence can also feed wider operational workflows through dashboards and APIs. Teams using SIEM or SOAR platforms may want to assess how a provider’s API and reporting fit their existing workflow rather than treating takedown as a separate queue. A provider should also record the evidence, actions, and outcomes behind each takedown, so the data supports reporting and feeds recurring attack patterns into wider security workflows.

Don't want to miss out on updates?

Don't want to miss out on updates?

Don't want to miss out on updates?

Join our mailing list for regular blog posts and case studies from Netcraft.