📈 Highlights

• Users' email addresses are exfiltrated to tracking, marketing and analytics domains before form submission and before giving consent on 1,844 websites when visited from the EU and 2,950 when visited from the US.
• We found incidental password collection on 52 websites by third-party session replay scripts. (These issues were fixed thanks to our disclosures).
• In a follow-up investigation, we found that Meta (formerly, Facebook) and TikTok collect hashed personal information from web forms even when the user does not submit the form and does not give consent.

Sample screen captures - Email address leaks

Sample screen captures - Leaks to Meta and TikTok due to Automatic Advanced Matching

⚠️ Leaks to Meta (Facebook) & TikTok

• Both Meta Pixel and TikTok Pixel has a feature called Automatic Advanced Matching [1, 2] that collects hashed personal identifiers from the web forms in an automated manner. The hashed personal identifiers are then used to target ads on the respective platforms, measure conversions, or create new custom audiences. (You can read about privacy issues caused by Meta Pixel on The Markup's Pixel Hunt series.)
• According to Meta's, and TikTok's documentation, Automatic Advanced Matching should trigger data collection when a user submits a form. We found that unlike what is claimed, both Meta and TikTok Pixel collects hashed personal data when the user clicks links or buttons that in no way resemble a submit button. In fact, Meta and TikTok scripts don't even try to recognize submit buttons, or listen to (form) submit events. You can view their overly broad and suspiciously similar list of selectors, which designates what page elements will trigger data collection. That means Meta and TikTok Pixel collect hashed personal information, even when a user decides to abandon a form, and clicks a button/link to navigate away from the page.
• In March 2022, we ran additional crawls of top 100K websites to detect leaks triggered by unrelated button or link clicks. Our crawler filled the email and password fields, and then clicked on a no-op (non-functional) button that it injected into the page. Injecting and clicking on the no-op button enabled us to detect leaks that would be triggered by Meta and TikTok's Automatic Advanced Matching. We found that 8,438 (US) / 7,379 (EU) sites may leak to Meta when the user clicks on virtually any button or a link, after filling up a form. In addition, we found 154 (US) / 147 (EU) sites that may leak to TikTok in a similar manner.
US Meta & TikTok leaks EU Meta & TikTok leaks
Email address leaks to tracker domains in the US-desktop crawl, no-action mode (no-action: no interaction with the consent dialog).



1. Site rank: Tranco rank
2. Encoding: Encoding or hash algorithm used when sending the email
3. Website: Hostname of the initially visited website (before a potential redirection)
4. Request Domain: eTLD+1 of the leaky request URL
5. Page URL: (last_page) URL of the page where our crawler filled the email field
6. XPath: XPath of the filled email field
7. Id:ID of the email element that our crawler filled
(5, 6 & 7 identify the leaking page and input elements. They can be used for debugging, reproduction etc.)

Email address leaks to tracker domains in the EU-desktop crawl, no-action mode (no-action: no interaction with the consent dialog).



1. Site rank: Tranco rank
2. Encoding: Encoding or hash algorithm used when sending the email
3. Website: Hostname of the initially visited website (before a potential redirection)
4. Request Domain: eTLD+1 of the leaky request URL
5. Page URL: (last_page) URL of the page where our crawler filled the email field
6. XPath: XPath of the filled email field
7. Id:ID of the email element that our crawler filled
(5, 6 & 7 identify the leaking page and input elements. They can be used for debugging, reproduction etc.)



• Full scale of leaks to Meta, and all leaks to TikTok were discovered after finalizing our paper, through the crawls described in the previous bullet point. There are two main reasons for that: 1) TikTok and Meta leaks are slightly different than the rest of the leaks we studied in our paper--they require further interaction with the page after filling out a form. 2) TikTok started beta testing the Automatic Advanced Matching feature in February 2022, long after we finish the crawls used in the paper (May'21). That also means, unlike the other results in our paper, some of the findings presented above (on Meta & TikTok pixel) are not peer-reviewed.
• We filed a bug report with Meta (25 March 2022), and reached out to TikTok using their Contact the Data Protection Officer form and Request privacy information form (21 April 2022). Meta swiftly responded to our bug report and said they assigned the issue to their engineering team (25 March 2022). TikTok hasn't yet responded to our disclosure. (The report to TikTok was filed more recently, since leaks to TikTok were discovered more recently--in fact while investigating the leaks to Meta.)

• Disclosure to Meta

  SubscribedButtonClick event fires on virtually every click, causing PII collection against user intent [Link]

  When Automatic Advanced Matching is enabled, SubscribedButtonClick event is fired after clicking virtually any button or link on a page. That means Meta Pixel collects hashed personal information, even when a user decides to abandon a form, and clicks a button/link to navigate away from the page.

  According to its official page [1], Automatic Advanced Matching should trigger data collection when a user submits a form: "After the visitor clicks Submit, the pixel's JavaScript code automatically detects and passes the relevant form fields to Facebook."

  Unlike what is claimed, Meta Pixel collects hashed personal data when the user clicks links or buttons that in no way resemble a submit button (attached screenshot). In fact, Meta's JavaScript code in question doesn't even try to recognize submit buttons, or listen to (form) submit events. See the attached screen captures:

  abcmouse.com (a website for children): Meta Pixel collects the hashed email address when the user closes the newsletter dialog. In that case, sharing the email address is the exact opposite of the user's intent ( attached screen capture).

  prothomalo.com: clicking Back, Terms of Service or Privacy Policy links triggers the collection of the hashed email address, and (hashed) first and last names. (attached screen capture)

  We hope you will recognize the disconnect between the described and actual behavior of Automatic Advanced Matching, and take necessary actions to address this issue.

  [1] https://web.archive.org/web/20220325001706/https://www.facebook.com/business/m/signalshealth/optimize/automatic-advanced-matching

  PS: This bug report is based on an academic study. If you need more technical details or have difficulties reproducing, contact [email address].

  Disclosure to Tiktok

  A slightly modified version of the following text is submitted to conform the word limit of the contact form.

  Automatic Advanced Matching causes PII collection against user intent

  When Automatic Advanced Matching is enabled, users' hashed email address is collected after clicking virtually any button on the page. That means TikTok Pixel collects hashed personal information, even when a user decides to abandon a form, and clicks a button or link to navigate away from the page.

  According to TikTok’s help pages [1], Automatic Advanced Matching should trigger data collection when a user submits a form:

  "Automatic Advanced Matching is programmed to recognize form fields, specifically capturing email and phone data. The capability is only triggered when the visitor submits information in an email or phone input field."

  Unlike what is claimed in the linked help page, TikTok Pixel also sends hashed personal data when the user clicks links or buttons that in no way resemble a submit button. In fact, the TikTok's JavaScript code in question doesn't even try to recognize submit buttons, or listens to form "submit" events.

  That means clicking almost any button or a link will trigger the data collection, regardless of the user’s intention. The issue is demonstrated in the following screen captures.

  Dismissing the consent dialog on kiwico.com/login causes Tiktok to collect the hashed email address from the login form: screen capture

  Clicking a radio button triggers collection on redcross.ca: screen capture

  On benjerry.com, hashed email collection is triggered by clicking on the email input field--not even a button or a link: screen capture

  We hope you will recognize the disconnect between the described [1] and actual behavior of Automatic Advanced Matching, and take necessary actions to address this issue.

  [1] https://archive.ph/BQ7nt#selection-3501.113-3501.217

  PS: This bug report is a result of an academic study. If you need more technical details or have difficulties reproducing, please reach us at: [contact email]

📣 Security Disclosures, GDPR Requests, and Leak Notifications

Our methods allow us to detect email and password leaks from clients to trackers, but what happens after the leaks reach third party's servers is unknown to us. In order to better understand the server-side processing of collected emails, and to disclose cases of password collection, we have reached out to more than a hundred first and third parties. In all cases we sent the emails using one of the authors' university email addresses, and real name; while also disclosing the nature of the research.

⏩ Follow-up Crawls

We ran additional crawls between 25-31 January 2022 to collect fresh data about the behavior we studied. Here are some highlights:

• Many websites where we detected leaks to Taboola started to use modal consent banners, which prevents interaction with the pages before giving consent. We didn't detect any email leaks on these websites without interacting with the consent banners, which substantially reduced the number of leaks to Taboola.
• Adroll started showing a consent dialog which reduced the number of leaks to Adroll to zero in the crawl. However, in manual follow up analysis we found several websites where Adroll collected hashed emails when the user simply clicked on the page. We have shared these examples with Adroll, who then addressed the issues and said that their "work to troubleshoot and deploy a comprehensive solution is underway with testing and incremental roll out".
• We identified incidental password collection by FullStory, Hotjar, Decibel and Yandex. Upon disclosures, Fullstory and Hotjar swiftly fixed the issues, which were due to mistakes on part of the first parties (e.g. copying the password values into other DOM elements' attributes). Yandex said they need some time to solve the issue, which they eventually did.
• In a manual follow-up investigation, we found additional password leaks to LogRocket on the login form of the zoning.sandiego.gov website. We've disclosed the issue to both LogRocket, zoning.sandiego.gov and opencounter.com (the provider of the web application running on zoning.sandiego.gov). We didn't get any response from those parties, but we verified that password leaks to LogRocket were eventually addressed.