Being a bad guy on the Internet is a really good business. In more than 90% of cybersecurity incidents, phishing is the root cause of the attack, and during this third week of August phishing attacks were reported against the U.S. elections, in the geopolitical conflict between the U.S., Israel, and Iran, and to cause $60M in corporate losses.

You might think that after 30 years of email being the top vector for attack and risk we are helpless to do anything about it, but that would be giving too much credit to bad actors, and a misunderstanding of how defenders focused on detections can take control and win. 

Phishing isn’t about email exclusively, or any specific protocol for that matter. Simply put, it is an attempt to get a person, like you or me, to take an action that unwittingly leads to damages. These attacks work because they appear to be authentic, visually or organizationally, such as pretending to be the CEO or CFO of your company, and when you break it down they are three main attack vectors that Cloudflare has seen most impactful from the bad emails we protect our customers from: 1. Clicking links (deceptive links are 35.6% of threat indicators) 2. Downloading files or malware  (malicious attachments are 1.9% of threat indicators) 3. Business email compromise (BEC) phishing that elicits money or intellectual property with no links or files (0.5% of threat indicators).

Today, we at Cloudflare see an increase in what we’ve termed multi-channel phishing. What other channels are there to send links, files and elicit BEC actions? There’s SMS (text messaging) and public and private messaging applications, which are increasingly common attack vectors that take advantage of the ability to send links over those channels, and also how people consume information and work. There’s cloud collaboration, where attackers rely on links, files, and BEC phishing on commonly used collaboration tools like Google Workspace, Atlassian, and Microsoft Office 365. And finally, there’s web and social phishing targeting people on LinkedIn and X. Ultimately, any attempt to stop phishing needs to be comprehensive enough to detect and protect against these different vectors.

_{Learn more about these technologies and products here}

It’s one thing to tell you this, but we’d love to give you an example of how a multi-channel phish plays out with a sophisticated attacker.

Here’s an email message that an executive notices is in their junk folder. That’s because our Email Security product noticed there’s something off about it and moved it there, but it relates to a project the executive is working on, so the executive thinks it’s legitimate. There’s a request for a company org chart, and the attacker knows that this is the kind of thing that’s going to be caught if they continue on email, so they include a link to a real Google form:

• The executive clicks the link, and because it is a legitimate Google form, it displays the following:

• There’s a request to upload the org chart here, and that’s what they try to do:

• The executive drags it in, but it doesn’t finish uploading because in the document there is an “internal only” watermark that our Gateway and digital loss prevention (DLP) engine detected, which in turn prevented the upload.

• Sophisticated attackers use urgency to drive better outcomes. Here, the attackers know the executive has an upcoming deadline for the consultant to report back to the CEO. Unable to upload the document, they respond back to the attacker. The attacker suggests that they try another method of upload or, in the worst case scenario, send the document on WhatsApp. 

• The executive attempts to upload the org chart to the website they were provided in the second email, not knowing that this site would have loaded malware, but because it was loaded in Cloudflare’s Browser Isolation, it kept the executive’s device safe. Most importantly, when trying to upload sensitive company documents, the action is stopped again:

• Finally they try WhatsApp, and again, we block it:

Setting up a security solution and maintaining it is critical to long term protection. However, having IT administration teams constantly tweak each product, configuration, and monitor each users’ needs is not only costly but risky as well, as it puts a large amount of overhead on these teams. 

Protecting the executive in the example above required just four steps:

1. Install and login to Cloudflare’s device agent for protection

   

   With just a few clicks, anyone with the device agent client can be protected against multi-channel phish, making it easy for end users and administrators. For organizations that don’t allow clients to be installed, an agentless deployment is also available.  

2. Configure policies that apply to all your user traffic routed through our secure web gateway. These policies can block access outright to high risk sites, such as those known to participate in phishing campaigns. For sites that may be suspicious, such as newly registered domains, isolated browser access allows users to access the website, but limits their interaction.

   

   The executive was also unable to upload the org chart to a free cloud storage service because their organization is using Cloudflare One’s Gateway and Browser Isolation solutions that were configured to load any free cloud storage websites in a remote isolated environment, which not only prevented the upload but also removed the ability to copy and paste information as well. Also, while the executive was able to converse with the bad actor over WhatsApp, their files were blocked because of Cloudflare One’s Gateway solution, configured by the administrator to block all uploads and downloads on WhatsApp. 

3. Set up DLP policies based on what shouldn’t be uploaded, typed, or copied and pasted.

   

   The executive was unable to upload the org chart to the Google form because the organization is using Cloudflare One’s Gateway and DLP solutions. This protection is implemented by configuring Gateway to block any DLP infraction, even on a valid website like Google.

4. Deploy Email Security and set up auto-move rules based on the types of emails detected.

   

In the example above, the executive never received any of the multiple malicious emails that were sent to them because Cloudflare’s Email Security was protecting their inbox. The phishing emails that did arrive were put into their Junk folder because the email was impersonating someone that didn’t match the signature in the email, and the configuration in Email Security automatically moved it there because of a one-click configuration set by the executive’s IT administrator.

But even with best-in-class detections, it goes without saying that it is important to have the ability to drill down on any metric to learn about individual users that are being impacted by an ongoing attack. Below is a mockup of our upcoming improved email security monitoring dashboard.

While phishing, despite being around for three decades, continues to be a clear and present danger, effective detections in a seamless and comprehensive solution are really the only way to stay protected these days. 

If you’re simply thinking about purchasing email security by itself, you can see why that just isn’t enough. Multi-layered protection is absolutely necessary to protect modern workforces, because work and data don’t just sit in email. They’re everywhere and on every device. Your phishing protection needs to be as well.

While you can do this by stitching together multiple vendors, it just won’t all work together. And besides the cost, a multi-vendor approach also usually increases overhead for investigation, maintenance, and uniformity for IT teams that are already stretched thin.

Whether or not you are at the start of your journey with Cloudflare, you can see how getting different parts of the Cloudflare One product suite can help holistically with phishing. And if you are already deep in your journey with Cloudflare, and are looking for 99.99% effective email detections trusted by the Fortune 500, global organizations, and even government entities, you can see how our Email Security helps. 

If you’re running Office 365, and you’d like to see what we can catch that your current provider cannot, you can start right now with Retro Scan.

And if you are using our Email Security solution already, you can learn more about our comprehensive protection here.

In the ever-evolving landscape of cyber threats, a subtle yet potent form of phishing has emerged — quishing, short for QR phishing. It has been 30 years since the invention of QR codes, yet quishing still poses a significant risk, especially after the era of COVID, when QR codes became the norm to check statuses, register for events, and even order food.

Since 2020, Cloudflare’s cloud email security solution (previously known as Area 1) has been at the forefront of fighting against quishing attacks, taking a proactive stance in dissecting them to better protect our customers. Let’s delve into the mechanisms behind QR phishing, explore why QR codes are a preferred tool for attackers, and review how Cloudflare contributes to the fight against this evolving threat.

The impact of phishing and quishing are quite similar, as both can result in users having their credentials compromised, devices compromised, or even financial loss. They also leverage malicious attachments or websites to provide bad actors the ability to access something they normally wouldn’t be able to. Where they differ is that quishing is typically highly targeted and uses a QR code to further obfuscate itself from detection.

Since phish detection engines require inputs like URLs or attachments inside an email in order to detect, quish succeeds by hampering the detection of these inputs. In Example A below, the phish’s URL was crawled and after two redirects landed on a malicious website that automatically tries to run key logging malware that copies login names and passwords. For Example A, this clearly sets off the detectors, but Example B has no link to crawl and therefore the same detections that worked on Example A are rendered inert.

Strange you say, if my phone can scan that QR code then can’t a detection engine recognize the QR code as well? Simply put, no, because phish detection engines are optimized for catching phish, but to identify and scan QR codes requires a completely different engine – a computer vision engine. This brings us to why QR codes are a preferred tool for attackers.

There are three main reasons QR codes are popular in phishing attacks. First, QR codes boast strong error correction capabilities, allowing them to withstand resizing, pixel shifting, variations in lighting, partial cropping, and other distortions. Indeed, computer vision models can scan QR codes, but identifying which section of an email, image, or webpage linked in an email has a QR code is quite difficult for a machine, and even more so if the QR codes have been obfuscated to hide themselves from some computer vision models. For example, by inverting them, blending them with other colors or images, or making them extremely small, computer vision models will have trouble even identifying the presence of QR codes, much less even being able to scan them. Though filters and additional processing can be applied to any image, not knowing what or where to apply makes the deobfuscation of a QR code an extremely expensive computational problem. This not only makes catching all quish hard, but is likely to cause frustration for an end user who won’t get their emails quickly because an image or blob of text looks similar to a QR code, resulting in delivery delays.

Even though computer vision models may have difficulty deobfuscating QR codes, we have discovered from experience that when a human encounters these obfuscated QR codes, with enough time and effort, they are usually able to scan the QR code. By doing everything from increasing the brightness of their screen, to printing out the email, to resizing the codes themselves, they can make a QR code that has been hidden from machines scan successfully.

Don’t believe us? Try it for yourself with the QR codes that have been obfuscated for machines. They all link to https://blog.cloudflare.com/

(Brick wall image by rawpixel.com on Freepik)

If you scanned any of the example QR codes above, you have just proven the next reason bad actors favor quish. The devices used for accessing QR codes are typically personal devices with a limited security posture, making them susceptible to exploitation. While secured corporate devices typically have measures to warn, stop, or sandbox users when they access malicious links, these protections are not available natively on personal devices. This can be especially worrisome, as we have seen a trend towards custom QR codes targeting executives in organizations.

QR codes can also be seamlessly layered in with other obfuscation techniques, such as encrypted attachments, mirrors that mimic well-known websites, validations to prove you are human before malicious content is revealed, and more. This versatility makes them an attractive choice for cybercriminals seeking innovative ways to deceive unsuspecting users by adding QR codes to previously successful phishing vectors that have now been blocked by security products.

Cloudflare has been at the forefront of defending against quishing attacks. We employ a multi-faceted approach, and instead of focusing on archaic, layered email configuration rules, we have trained our machine learning (ML) detection models on almost a decade’s worth of detection data and have a swath of proactive computer vision models to ensure all of our customers start with a turnkey solution.

For quish detections, we break it into two parts: 1) identification and scanning of QR codes 2) analysis of decoded QR codes.

The first part is solved by our own QR code detection heuristics that inform how, when, and where for our computer vision models to execute. We then leverage the newest libraries and tools to help identify, process, and most importantly decode QR codes. While it is relatively easy for a human to identify a QR code, there is almost no limit to how many ways they can be obfuscated to machines. The examples we provided above are just a small sample of what we’ve seen in the wild, and bad actors are constantly discovering new methods to make QR codes hard to quickly find and identify, making it a constant cat and mouse game that requires us to regularly update our tools for the trending obfuscation technique.

The second part, analysis of decoded QR codes, goes through all the same treatment we apply to phish and then some. We have engines that deconstruct complex URLs and drill down to the final URL, from redirect to redirect, whether they are automatic or not. Along the way, we scan for malicious attachments and malicious websites and log findings for future detections to cross-reference. If we encounter any files or content that are encrypted or password protected, we leverage another group of engines that attempt to decrypt and unprotect them, so we can identify if there was any obfuscated malicious content. Most importantly, with all of this information, we continuously update our databases with this new data, including the obfuscation of the QR code, to make better assessments of similar attacks that leverage the methods we have documented.

However, even with a well-trained suite of phish detection tools, quite often the malicious content is at the end of a long chain of redirects that prevent automated web crawlers from identifying anything at all, much less malicious content. In between redirects, there might be a hard block that requires human validation, such as a CAPTCHA, which makes it virtually impossible for an automated process to crawl past, and therefore unable to classify any content at all. Or there might be a conditional block with campaign identification requirements, so if anyone is outside the original target’s region or has a web browser and operating system version that doesn’t meet the campaign requirements, they would simply view a benign website, while the target would be exposed to the malicious content. Over the years, we have built tools to identify and pass these validations, so we can determine malicious content that may be there.

However, even with all the technologies we’ve built over the years, there are cases where we aren’t able to easily get to the final content. In those cases, our link reputation machine learning models, which have been trained on multiple years of scanned links and their metadata, have proven to be quite valuable and are easily applied after QR codes are decoded as well. By correlating things like domain metadata, URL structure, URL query strings, and our own historical data sets, we are able to make inferences to protect our customers. We also take a proactive approach and leverage our ML models to tell us where to hunt for QR codes, even if they aren’t immediately obvious, and by scrutinizing domains, sentiment, context, IP addresses, historical use, and social patterns between senders and recipients, Cloudflare identifies and neutralizes potential threats before they can inflict harm.

With the thousands of QR codes we process daily, we see some interesting trends. Notable companies, including Microsoft and DocuSign, have frequently been the subjects of impersonation for quishing attacks. What makes this more confusing for users, and even more likely to scam them, is that these companies actually use QR codes in their legitimate workflows. This further underscores the urgency for organizations to fortify their defenses against this evolving threat.

Below are three examples of the most interesting quish we have found and compared against the real use cases by the respective companies. The QR codes used in these emails have been masked.

Microsoft uses QR codes as a faster way to complete MFA instead of sending six digit SMS codes to users’ phones that can be delayed and are also considered safer, as SMS MFA can be intercepted through SIM swap attacks. Users would have independently registered their devices and would have previously seen the registration screen on the right, so receiving an email that says they need to re-authenticate doesn’t seem especially odd.

DocuSign uses QR codes to make it easier for users to download their mobile app tosign documents, identity verification via a mobile device to take photos, and supports embedding DocuSign features in third party apps which have their own QR code scanning functionality. The use of QR codes in native DocuSign apps and non-native apps makes it confusing for frequent DocuSign users and not at all peculiar for users that rarely use DocuSign. While the QR code for downloading the DocuSign app is not used in signature requests, to a frequent user, it might just seem like a fast method to open the request in the app they already have downloaded on their mobile device.

Microsoft uses QR codes for Teams to allow users to quickly join a team via a mobile device, and while Teams doesn’t use QR codes for voicemails, it does have a voicemail feature. The email on the left seems like a reminder to check voicemail in Teams and combines the two real use cases on the right.

As we confront the persistent threat of quishing, it's crucial for individuals and organizations to be vigilant.  While no solution can guarantee 100% protection, collective diligence can significantly reduce the risk, and we encourage collaboration in the fight against quishing.

If you are already a Cloud Email Security customer, we remind you to submit instances of quish from within our portal to help stop current threats and enhance the capabilities of future machine learning models, leading to more proactive defense strategies. If you aren’t a customer, you can still submit original quish samples as an attachment in EML format to quish@cloudflare.com, and remember to leverage your email security provider’s submission process to inform them of these quishing vectors as well.

The battle against quishing is ongoing, requiring continuous innovation and collaboration. To support submissions of quish, we are developing new methods for customers to provide targeted feedback to our models and also adding additional transparency to our metrics to facilitate tracking a variety of vectors, including quish.