Security
What Bloom sees, and how it's protected
Bloom runs in your site's pages, so it sees what your users see. Here's what it reads and what it keeps out, where that goes, who can open it and how it's protected. It describes what Bloom's code does today.
Updated November 19, 2026
Sensitive fields are masked in the browser
Fields that look like passwords, card and account numbers or one-time codes are masked before anything leaves the visitor's device.
You choose what Bloom can see
You can exclude private content from page reads and recordings, mask text, or reserve a visible control for the visitor.
It works with the visitor's own access
Bloom acts in the visitor's browser, signed in as they are, and is told to act only when they ask. It has no access of its own to your servers or database.
Encrypted in transit and at rest
HTTPS to Bloom, TLS to its database, and AES-256 on the database, its backups and recordings.
Dashboard access is scoped to your site
The dashboard checks team membership and looks up records within the site being opened. Automated tests cover access across sites.
Off by default in production
With the tag from your Install page, Bloom stays off in production until you turn it on, so you can try it in staging first. Each environment has its own switch.
Compliance
| Status | What it means | |
|---|---|---|
| SOC 2 | Not yet | Bloom hasn't completed a SOC 2 audit. Our team can answer your security questionnaire. |
| Our hosting providers | Audited | AWS and Cloudflare, which Bloom runs on, have SOC 2 reports of their own, covering their services. |
| Data protection terms | In our Terms | Section C.4 of our Terms of Service, under which the people who process personal data for Bloom are bound to confidentiality. |
How data moves
Your visitor's browser
Bloom's widget reads the page and records the visit
Stays in the browser Masked field values and text, content excluded from recording, and the contents of other sites' frames
HTTPS, through Cloudflare
Bloom, on AWS in Oregon, US
Bloom's servers
Database and recordings, encrypted
The visit's page text, chat, replay and visitor details
AI providers
The models that write Bloom's answers, summaries and reports
Google sign-in, for your site's team
Your dashboard
Sessions, replays, users and tickets
Multi-tenancy and data isolation
Bloom is multi-tenant: every customer's site runs on one service, with one database and one file storage, and each site's data is kept apart in software. This is what keeps it apart.
Your site's visitors
Another customer's visitors
Your site's key
Their site's key
Bloom: one service, one database, one file storage
Your site's sessions, recordings, users, tickets, settings and keys
Their site's sessions, recordings, users, tickets, settings and keys
Kept apart by site
Google sign-in, and your Team page
Google sign-in, and their Team page
Your team's dashboard
Their team's dashboard
Every record belongs to one site
Each site's sessions, recordings, users, tickets, emails and settings carry its ID, or belong to a session or ticket that does.
How the dashboard checks access
For customer access, the dashboard checks the Google-verified email against the site’s Team page. Record lookups are scoped to that site. Who can open it also describes access for authorized Bloom staff.
Tested on every change
On every change, automated tests try each dashboard address whose path names a record, and each of the MCP server's tools, with another site's records.
A visitor reaches only their own session
The widget sends data in. It has no request that reads a site's sessions, recordings or tickets back: those are only in the dashboard. What a visitor's browser can get is its own chat, by its session's random ID. The key in your script tag isn't a secret: it identifies your site, and gets only what any visitor's widget gets.
AI requests are made for one site
Each request to an AI model is built from one site's own instructions, skills, map and help center, and that site's sessions: one session's page and chat for an answer, the week's sessions for Insights. So one customer's data doesn't show up in another's answers or reports.
Your keys are tied to your site
Credentials you give Bloom for another service, like Zendesk or Plain, are encrypted and tied to your site, so one site's can't be used as another's.
Email is per site
Email reaches Bloom at an address that's your site's alone, and your team's replies go out from your own domain.
What Bloom reads and records
Wherever the widget is on and showing to a visitor, from when the page loads:
The page, to help them
When the visitor asks for something, after each thing Bloom does, and before it offers help, the widget reads the page as text: what's on it, and its links, buttons and fields, with what's in them.
The visit, to replay it
It records every visit, whether or not the visitor chats: the page as it changes, scrolling, mouse moves, clicks and what's typed into fields.
The page's own requests
It watches the page's fetch and XMLHttpRequest calls without changing them, to know when the page has settled and whether a save failed. Of requests to your own site, it keeps the method, the path and the status, never what they send or get back.
The chat
What the visitor wrote, Bloom's replies, what it did, and any rating the visitor gave a reply.
Who the visitor is, if you say
What your page passes to Bloom.updateUser, like a user ID, name and email. Anything you pass as privateAttributes is dropped in the browser.
Their device and rough location
The browser and operating system, and the city, region and country their IP address is in. A session keeps the place, not the IP address.
On the visitor's device
The widget keeps the chat in the tab's session storage, and in local storage a random ID for the browser and the visitor's choices, like hiding Bloom.
What never leaves the browser
Masking happens in the visitor's browser, before anything is sent, so what's kept out never reaches Bloom's servers or an AI model.
Fields that look sensitive
Passwords, card and account numbers, one-time codes and the like, going by each field's type, autocomplete hint, name, ID, placeholder and labels. Bloom doesn't read what's in them, and the replay shows asterisks. It's a best guess.
What you mark
<p class="rr-block">Card ending 4242</p>
<p>Signed in as <span class="rr-mask">Alice Moreno</span></p>
<button data-bloom-block>Pay $42</button>Use rr-block to exclude an element from page reads and recordings. Use rr-mask for text and field values. To exclude private attributes or a shadow root, apply rr-block to the element or its shadow host. Mark private content inside same-origin frames directly.
A control marked only with data-bloom-block stays visible to Bloom and in replays. The visitor controls its use.
Frames from other sites
Like a payment provider's card form. Bloom doesn't read inside them.
Pages and visitors Bloom doesn't show for
Where your rules keep Bloom off a page or away from a visitor, or a visitor has hidden it, the page isn't read or recorded, though the widget still checks in with Bloom's server, with who your page says the visitor is. A visitor already in a conversation keeps Bloom on a blocked page until they leave it.
Everything else on a page is read and recorded as it is, email addresses included, so mark what's private.
What the agent can do
What the visitor asks for
Bloom is told to change a page only when the visitor asked it to, or said yes when it offered, and to answer a question rather than act. The visitor can stop it at any time, and the controls you reserve are enforced in code.
Only what the visitor could do
It clicks and types in the page, in the visitor's browser, signed in as they are, with no access of its own to your servers or database. So your server's checks, like who may pay or delete, hold for everything Bloom does on your site.
Never the controls you reserve
Mark a control, a form or a part of a page with data-bloom-block, like a button that pays, and Bloom won't click, type into or submit it. It points it out for the visitor to press.
The web
For what your site doesn't say, Bloom can search the web, through its AI provider's search tool, and read public pages from its servers, which check that an address is on the public web before they fetch it.
AI models
Bloom's answers, summaries and reports are written by AI models, which are other companies', so what's sent to one leaves Bloom's servers for that company's. Which companies, and what each does, is below and on our Subprocessors page.
| What the model does | What it gets |
|---|---|
| Answering a visitor | The page as the widget read it, the chat, and what your page told Bloom about the visitor |
| Summarizing a visit, once it's over | A video of its replay, its chat, and a list of what happened in it |
| Writing the week's Insights | The week's sessions, with their summaries, chats and ratings, each visitor's email where your page gave one, and their ticket numbers |
| Running and grading your Tests | The moments each test keeps |
| Finding help articles | The searches Bloom writes, and your help center's articles |
| Checking on a visitor who seems stuck, where you've turned that on | The page as the widget read it, and the chat so far |
What Bloom may train on
Our Terms let Bloom train its AI features on what goes to and comes from them only once it's de-identified, so it identifies no one, and aggregated with data across other customers (Terms, section D.3).
Where data is kept, and how
In the United States
Bloom's servers and database are on AWS in Oregon (us-west-2). Cloudflare, in front of all of Bloom's addresses, handles requests at its edge, around the world.
Encrypted in transit
Everything between a browser and Bloom travels over HTTPS, and Bloom's servers connect to their database over TLS.
Encrypted at rest
The database, its backups and the files Bloom keeps, like recordings and email attachments, are encrypted with AES-256 by AWS, in storage that isn't public.
The database
Postgres on Amazon RDS, with no direct route from the internet, and automated backups kept for 7 days.
Who can open it
Your team
Everyone on your site's Team page sees all of it, its sessions, replays, users and tickets, and can change its setup.
Visitors
A visitor sees only their own chat.
Bloom's team
Authorized Bloom staff can open your site in the dashboard and access production data, to support you and run the service.
Signing in
With Google, so your Google account's protections, like 2-Step Verification, apply. You're signed out 30 days after you last opened the dashboard.
A coding agent setting Bloom up
The token it gets works only for that one site's setup, once a signed-in person approves it, and lasts a day. It can't read sessions, recordings or tickets.
GitHub, if you connect it
Bloom's GitHub App can only read, and only the repositories you pick. Bloom doesn't keep your GitHub token.
Claude Code and Codex
Through Bloom's MCP server, they read only what you can see in the dashboard, and change nothing. Only apps on your own computer can sign in, and taking someone off your Team page ends what their apps can read.
How we build and ship Bloom
Tested on every pull request
Every pull request runs Bloom's tests, including browser tests that check masked and blocked content stays out of what Bloom reads and records.
One source for the widget
Everything the widget runs is served from js.bloomresearch.ai, including Bloom's own copies of the libraries it uses.
Least privilege
Each of Bloom's services has only the permissions it needs: the web servers can read recordings, but not change or delete them.
What you control
Environments
With the tag from your Install page, Bloom is off in production until someone on your team turns it on, while staging and localhost work as soon as the tag is there. Each environment has its own switch, on Configure.
Pages
Choose the pages Bloom shows on, and block the ones it shouldn't.
Visitors
Choose who sees Bloom, by email. The list stays on Bloom's server, so visitors can't read it.
Your team
Add and remove people on the Team page, which shows when each was last active.
Your visitors' choice
Any visitor can hide Bloom from its menu, for a day or for good in that browser, and it stops recording at once.
A history of changes
Bloom keeps each revision of your site's configuration, like its instructions, skills, page rules, visitor settings and nudges, with who made it and when.
Taking it out
Remove the script tag and the Bloom.updateUser calls. Without the widget, the marks do nothing, so they can stay.
Who processes data for Bloom
These companies process data for Bloom, as our Subprocessors page lists them.
| Company | What it does for Bloom | Location |
|---|---|---|
| Amazon Web Services | Hosting, database, storage of session recordings, email | United States |
| Cloudflare | Delivering the widget, security, approximate location of visitors | Worldwide |
| Anthropic | AI models for Bloom's agent and reports | United States |
| AI models for session summaries; sign-in with Google | United States | |
| OpenAI | AI models for search over help articles | United States |
| TypeSafe | Signs that a visitor may be stuck, from page text | United States |
| Sentry | Error tracking | United States |
| PostHog | Product analytics and experiments | United States |
And only when you connect one of them: Zendesk, for support tickets; Plain, for support tickets; Crisp, for handing visitors over to a person; GitHub, for reading a site's code to map it.
Keeping and deleting data
Deletion
Sessions, recordings, and tickets don't expire automatically. Contact Bloom to request deletion.
Logs
Bloom's logs are kept for 30 days.
Where you connect your support address, Bloom keeps each email forwarded to it, with its attachments, and each reply your team sends from Support.
Reporting a vulnerability
If you think you've found a security problem in Bloom, tell us privately first, through Contact at the bottom of this page, with what you found and how to reproduce it. Please don't access other people's data, or disrupt Bloom, while you look.