Security
Last updated 16 November 2026
Peoplume holds salaries, bank details and identity documents, so this page describes how it is built rather than reassuring you that it is secure. The last section lists what it does not do, because that is usually the part you are trying to find out.
One company cannot see another
Every record in the database carries the company it belongs to, and every query filters on it. There is no shared pool of employees that a forgotten filter could expose, and company scope alone is not treated as sufficient — reads are also checked against the relationship that entitles you to them: your own record, your own direct report, a leave request your leave type names you as the second approver of (as the head of that person’s department, or as their manager’s manager), a timesheet (with its attended hours) of someone in a department you head when your company has department heads approve timesheets, your own document.
Identifiers arriving in a request are never trusted as targets. A department, branch, manager or role referenced by a form is looked up within your company before anything is written, so an identifier belonging to another customer simply does not resolve.
Passwords and sessions
- Passwords are hashed with bcrypt. They are never stored or logged in readable form. Every new password — chosen at sign-up, from an invitation or a reset link, when changing one’s own, or typed by an administrator — needs at least 8 characters and is refused when it is one of the 10,000 most common passwords or contains the email address’s name, the company’s name or “Peoplume”.
- The session cookie is HTTP-only, same-site, and marked secure in production. It carries only a user identifier — role, company, permissions and branch scope are re-derived from the database on every single request, so a role change takes effect immediately and no privilege can be smuggled inside a token.
- Changing or resetting a password signs out every other browser and phone, and a sign-in still waiting for its two-step code, as does “sign out everywhere”, which everyone has, portal users included. Each of these also turns off push notifications on all of the account’s browsers and phones. Tokens issued before that moment stop being accepted. Someone who may manage users can sign another person out of every browser and phone the same way, if they hold every permission that person holds (for an Administrator, only another Administrator).
- When email is set up and the person’s address has been shown to be theirs (by a link from it they opened, or a password they chose themselves), an administrator’s password reset emails them a link that works for an hour instead of setting a password, and their current password keeps working until the link is used; otherwise the administrator types a temporary password, which the person must replace when they next sign in. People are told, in the app and by email (even with email copies turned off, once their address has been shown to be theirs), whenever someone else resets their password, sends them a reset link or clears their two-step verification, whenever their password changes, when a change of their sign-in address is asked for and when it is made, and when their portal login is given console access.
- Someone who may manage users can change a login’s name, or ask to change its sign-in address, if they hold every permission that account holds (for an Administrator, only another Administrator) — their own account included. A sign-in address changes only when the new address confirms it through an emailed link that lasts 24 hours, with the account’s own two-step code, or its password if it has no two-step verification — or, for an account required single sign-on covers that has none, while signed in as that account through the company sign-in within the last 15 minutes; never by the link alone. An account required single sign-on covers keeps an address on the company’s proven domains. A password change or signing out everywhere cancels a pending link, and a new request replaces it. The change signs the account out of every browser and phone and cancels any password-reset link sent to the old address; the person is told when it is asked for and, at their old address, when it is made, as are the company’s Administrators.
- Every sign-in is recorded in the company’s audit log with how it was made — a password, a two-step or recovery code, single sign-on, the phone app, accepting an invitation or confirming the sign-up — never with a network address or a device, so the log says when and how someone signed in, not from where. The log’s default view leaves sign-ins out; people allowed to view it see them under their own filter. Wrong passwords, and wrong two-step codes after a right password or a single sign-on, typed when signing in on the web or in the phone app, are counted on the account — a count goes on until 30 days pass without a failure, and a successful sign-in does not clear it — and shown, until then, to whoever may manage users and, for a console account, to its holder in Settings. An attempt refused by a limit is not counted.
- A browser sign-in lasts 7 days from when it was made. A company can shorten that for console (HR and administrator) accounts to 3 days, 24, 12 or 8 hours, which also ends older sign-ins; portal sign-ins and the phone app are not shortened.
- A leaver’s portal and app login works through their last day and is deleted the day after it — when the termination is saved if that day has already passed, otherwise by a daily sweep — which signs them out of every browser and phone at once.
- Accounts with console (HR or administrator) access are not deleted that way: their access ends on a date set on the account itself. Recording someone’s exit offers to end their console access on their last day — ticked whenever the person recording it may manage users and act on that account — and Settings can set, change or clear the date. The holder keeps the console through that date, judged on the company’s own calendar day; from the next day the account is suspended — any session it already has stops working, on the web and in the app, signing in is refused with a plain sentence saying access has ended, a password-reset request sends no link, and it stops receiving notifications, email copies, pushes and the monthly digest. It is kept rather than deleted, so the audit log still names it, and whoever may manage the account can restore it. Restoring a suspended account — like ending access on a day already past — signs it out of every browser and app device, so nothing it held before is revived: its holder signs in afresh. Only someone allowed to act on the account can set the date, and never on their own. The date is never derived from the employee record: moving a last day or reinstating someone leaves it as it is (when an end date is set, whoever manages the account is told to review it), so editing a record cannot cut anyone off. A company always keeps at least one Administrator whose access has no end date — the last one cannot be given an end date, moved to another role or removed. An account whose holder has left with no end date set is flagged to whoever manages users, in Settings and on their dashboard, and that person’s record cannot be erased or deleted while a console account tied to it (linked, or signing in with its email) still has access — the access is ended first, so the flag cannot disappear while the account still works. Ending an account does not revoke the API keys or webhooks it set up: those belong to the company, are usually wired into its own tools, and keep working until someone revokes them. So once an account’s access has ended, or its holder has left, Settings lists beside it the live keys and webhooks the audit log shows it created, for someone to review — nothing is revoked automatically.
- The Peoplume phone app signs in with the same password and the same two-step code, under the same rate limits. It receives a device token that we store only as a SHA-256 hash; it lapses after 60 days unused, can be signed out from the website, and stops working the moment the password changes or the account is signed out everywhere. The app reaches only the employee’s own records — and, for a manager, their direct reports’ leave requests; for a department head or a manager’s manager a leave type names as second approver, the leave requests waiting for their final approval. It uses the camera, the photo library and the phone’s files only to attach a receipt to the employee’s own expense claim, when they choose to; a photo is re-encoded on the phone before it is sent, and it reaches Anthropic only when “Read the receipt” is pressed while the employer has receipt reading switched on. Travel advances are read-only in the app. An employee can turn on an app lock that asks for the phone’s own fingerprint, face or passcode (the phone checks it; we never see it), and payslip screens are kept out of screenshots and the app-switcher preview. Crash reports from the app come to our own servers with email addresses, tokens and long numbers removed, and are logged, not stored.
- Password resets, invitations, links confirming a new sign-in address, interview invitations, offer response links and API keys are stored only as SHA-256 hashes. Reading the database yields no working link or key.
- A candidate’s offer link works once, stops at the end of the respond-by day on the employer’s clock and whenever the offer is changed, re-sent, withdrawn or decided by the employer, and records an answer only from a confirmed button press — never from opening the link, so a mail scanner records nothing. It keeps no IP address or device, and accepting through it is a record of the answer, not an electronic signature. A self-scheduling link a recruiter emails to a candidate is built from the configured application address, never from a request header, goes once per link, and neither the link nor the email’s text is stored.
- Signing in, on the web and in the phone app, is limited per account, and per network address on failed attempts and, more loosely, on every attempt; failed attempts against existing accounts of companies that are not in a free trial or read-only are also limited per email domain (other than webmail domains such as gmail.com) and across Peoplume, each account — and, across Peoplume, each company — counting only up to a share, so neither guesses at addresses that have no account nor one customer’s own accounts can use up the limit across Peoplume. Checking a password or a clock-in PIN, and storing a new one, also take their turn in a pace shared across Peoplume, one for each kind — signing in, clock-in presses, and every other password step — so a flood of one kind does not hold up the others, and each company’s other password steps count only up to a share of their pace; at busy times a check waits briefly (at most three seconds for signing in and clock-in presses, one for the other steps) or is asked to try again in a moment, and that answer never counts as a wrong password or PIN. Setting clock-in PINs, and typing a new password for someone else, are limited per person. Two-step codes are limited per account and per network address, and wrong codes (not those typed while first setting it up) also across Peoplume in the same way, whether they are entered at sign-in or inside Settings; password checks inside Settings are limited per account. Sign-up is limited per network address, per email address and across Peoplume; password-reset requests per network address and per email address, with reset emails capped across Peoplume; accepting an invitation per network address and across Peoplume. Adding users directly and sending invitations are limited per person and per company; someone who may manage users can send portal invitations to up to 500 people at once, only where email is set up — each the same single-use, seven-day link, sent from Peoplume, and one whose email fails is withdrawn. Asking to change a sign-in address is limited per person and per account, and confirming one per network address and across Peoplume. Starting a single sign-on and returning from the provider are limited per network address, and a provider’s published settings, once fetched and checked, are reused for five minutes of sign-ins. API requests are limited per key, and per network address before the key is even checked. Job applications, the clock-in kiosk, interview booking, offer responses and the careers feed and logo are limited per network address before their link is looked up. Network addresses are the ones Cloudflare reports, an IPv6 address counted with the rest of its /64 block; the server can be set to refuse requests that did not come through Cloudflare.
- A new company exists only after its administrator confirms their email address: the sign-up form stores a pending request, and the company, the account and the trial are created when the emailed link is confirmed — by a button, not on arrival, so a mail scanner opening the link creates nothing. The administrator chooses their password on that confirmation page, which shows the Terms of Service and the Data Processing Agreement beside the button and records which version was accepted, when and by whom.
Two-step verification
Any user can turn on TOTP two-step verification from their own settings, using any standard authenticator app. Setting it up asks for the account’s password — or, for an account its company’s Require single sign-on covers, which may have no password it knows, a sign-in with the company account within the last 15 minutes. Enrolment only takes effect once a code has been accepted, so an abandoned set-up leaves the account exactly as it was, and turning it off again requires a current code — a stolen session cannot quietly remove it. A code works once: one already accepted, at sign-in or in Settings, is refused if it is entered again.
Ten single-use recovery codes are issued once and stored hashed. If someone loses both their authenticator and their codes, someone who may manage users and holds every permission they hold (for an Administrator, another Administrator) can clear it; it is recorded in the audit log, the person is told, and nobody can clear their own. A company’s only Administrator, whom nobody inside can clear it for, can ask Peoplume support: after checking who they are by a separate channel, a Peoplume owner clears it with a command run outside the product, recorded in the company’s own audit log under Peoplume’s name; every session of the account is signed out, and the person and the company’s other Administrators are told. An only Administrator turning two-step verification on is told this, and advised to appoint a second Administrator.
An administrator can require it for everyone with console access — people who see salaries and run payroll. Portal accounts are excluded on purpose: they include staff who clock in at a shared tablet precisely because they have no smartphone, and a rule they cannot satisfy is a lockout rather than a control. Enabling the requirement is refused unless the administrator has already enrolled themselves.
Single sign-on
A company can connect its own identity provider over OpenID Connect — Microsoft Entra ID, Google Workspace, Okta and others. Only its Administrators can set it up or change it; every change but checking a domain’s DNS record asks them again for their password or two-step code (or, where passwords are switched off for them, a sign-in through the provider in the last 15 minutes), and every Administrator is told of every change.
It covers only email domains the company has proven it controls by publishing a DNS record we check — up to five; webmail domains such as gmail.com are refused, and a domain nobody has proven never blocks anyone.
The ID token is validated against the provider’s published keys; the flow uses PKCE with a state and nonce bound to the browser that started it, and starts only from our own sign-in page, which names the provider’s address before sending you there. Every address we fetch for it must be a public https address, and the request carrying the company’s client secret never follows a redirect.
Anyone who has turned on two-step verification is still asked for their code after the provider signs them in.
Passwords keep working alongside single sign-on unless the company turns on Require single sign-on: console accounts on its proven domains are then refused password sign-in on the web and in the phone app, and password-reset requests send them no link; sessions they already have open are signed out when it is turned on. One named Administrator keeps a password in case the provider fails, and portal accounts keep theirs. The phone app cannot sign in through a company’s provider yet.
Two limits worth knowing before you plan a rollout. Signing in this way requires an existing Peoplume account: we do not create users from an assertion, because that would let a misconfigured group mapping become a stranger with a login. And an address outside the email domains the company has proven is refused however well-signed it is.
Who can see what
Permissions are individual rather than a handful of fixed roles: viewing salaries is separate from editing them, and both are separate from running payroll. A role can be scoped to one branch, in which case its holder sees only that branch’s employees — enforced in the queries themselves, not by hiding menu items.
Roles are contained: you can only grant permissions you hold yourself, so an administrator cannot mint an account more powerful than their own, and only an Administrator can make someone an Administrator. An employee’s portal login becomes an HR console account only when someone who may manage users gives it a role whose every permission they hold themselves, and never on their own record; it keeps the same login, password, two-step set-up and history, the person is told, and a leaver serving notice keeps console access only to their last day. Salary changes, approvals, deletions and permission changes are written to an append-only audit log naming who did them; its pay figures show only to people who may view salaries, and a search of the log never matches them for anyone else. People allowed to view the log can filter it by date, person and action and download what they filtered as a spreadsheet; each download is itself recorded, and a download’s pay figures follow the same rule.
The full data archive needs permission to manage settings, and parts of it follow the permissions of the person downloading it: salaries, salary history and bank details need permission to view salaries (the full IBAN, permission to change them), payslips also permission to view payroll, the audit log permission to view it, and recruitment records permission to view recruitment. The uploaded-files download also needs permission to manage settings, and gives documents, receipts, CVs and training certificates only to someone with the permission of the page that shows them. Anything withheld is named in the download’s README.
Candidate access in the employee portal is by assignment to an opening (hiring manager or team member), not by role, and is checked on every request: it ends when the person is removed from the team, stops working here, or 90 days after the opening closes; nobody sees an opening they applied for. Each interviewer on a panel writes their own feedback in the employee portal and, signed in with the login linked to their employee record, sees other panelists’ feedback — there and on the candidate’s page in the console — only once they have submitted theirs; unsubmitted drafts are never shown to anyone else. That is a rule for writing independent feedback, not an access control: the downloads made for HR (a candidate’s data, an employee’s data and the company archive) hold submitted feedback whether or not the person downloading sits on the panel. Nobody marks their own employee record as a candidate, records themselves as a referrer or changes a referral recorded for them. Recruitment’s audit entries include candidate exports, candidate deletions (with counts, never the email), the applicants spreadsheet, internal-candidate links, referrer changes, put-forwards and internal moves (employee numbers, never email addresses), and one entry for a bulk stage move. Deleting a candidate is one transaction bound to the applications the page showed: an application that arrived meanwhile is never deleted unseen, and CV files are removed only after the deletion commits.
Engagement survey answers are kept apart from who answered. A response is stored with no name and no time, and the survey’s responses are rewritten in a random order after every answer, so their order in storage says nothing about who answered when; the record of who responded keeps only the day. Results appear only once the survey is closed and at least its minimum number of people have answered — while it runs, only the count is shown, so no single answer can be watched arriving — and a survey whose results were shown cannot be reopened. That holds for the product and for a copy of the database: each response keeps one storage key, fixed by its random identifier, through every rewrite, so anything the database engine leaves in unused space only repeats an answer under that same key, and that space is cleared every night; the space old rows used is overwritten when the responses are rewritten, and backups copy the current rows only. Two copies taken at different times (two backups, say, or a disk image, which can also hold the temporary journal the database writes during an answer) show which answers arrived between them and who answered in between, so someone holding the database’s backups could tell apart the answer of a person who was the only one to answer in that interval. While a survey is open its count of people who could still answer is kept current, and someone coming off the list of people who can answer (a leaver, or a login removed) lowers it only if they had not answered — so whether that person took part, never what they answered, can be read from it.
A survey’s extra statements are stored on the same terms, each answer on its own with nothing linking it to the person or to their other answers. An employer can choose, when creating a survey, to break its results down by branch or department, which narrows anonymity and is said so on the form: an answer is then stored with the respondent’s branch or department, never the person — but only for a group large enough to stand apart when the survey was created, with at least the survey’s minimum number of people who could answer it then; every other group’s answers (groups too small then, a larger group pooled beside them when they alone would be too few — or, when no group is too small, the smallest group anyway, so the pool is never empty — and any group created later) are pooled and stored with no group, and which groups are which does not change while the survey runs. A group’s results are shown only once the survey is closed and at least its minimum number of people in that group have answered, and groups are held back until those not shown together hold at least the survey’s minimum, so the total cannot give them away. A breakdown names only the groups it shows, then one line for all the others whatever was answered — a group not shown is never named, counted or given a figure — comments are never broken down, and the archive export applies the same rule. The stored group is what a breakdown adds to a copy of the database: whoever holds that copy can tie the answers stored under a group, the pooled ones included, to that group’s people who answered — to one person, when only one of them did, which a group large enough when the survey was created, the pool included, can still come to if its members leave or do not answer.
A manager’s private notes on a one-to-one are shown to that manager and to people allowed to manage performance — in the full data archive, only to an administrator who may also manage performance — and never reach the employee they are about, by any channel: no page, no notification, no app screen, and no download — their own names them as held and not given, and the archive’s one-to-ones file leaves out every one-to-one about whoever downloads it, while its audit log, which keeps every entry, records only that a one-to-one was recorded or edited, with names and the date. A manager’s note for HR on a probation recommendation is held from the employee the same way. In the archive, the suggested peer reviewers file leaves out suggestions for the downloader’s own review, and its audit log still records each suggestion, each decision on one and each nomination, with names and dates — never the reasons. The per-person export gives each sensitive part only to a reader holding the permission of the page that shows it: self-assessments and 360° feedback need permission to write reviews, not only to view results. It leaves out every row about the reader — 360° feedback a colleague gave about them, suggestions of peer reviewers for their own review or naming them. Someone exporting their own record from the console through an account linked to it gets the employee’s own download; a console user whose sign-in email is merely on a record gets only what their own permissions allow, never the parts the employee’s own download holds back, and that export is logged like anyone else’s.
What Peoplume’s own staff can see
Peoplume’s own staff run the service from a separate panel that shows companies, not the people in them: a company’s name; its counts of employees, branches, departments and accounts, and how many of its console accounts have two-step verification on; its subscription and the payment provider’s messages about it; when it last used the web or the app, the phone app, a kiosk and the API; its settings switches and the version of our terms it accepted; what its use of the AI features has cost today and this month, against the service’s daily ceiling and its own allowance; a deletion it asked for or withdrew, with what still stands in its way; and, for a company deleted, the record kept of it: its name and id, when, and how much was deleted. About the service itself, the panel shows whether each of its settings is in place — yes or no, never the value — and how old the newest database snapshot is. It never shows the name, email address, pay or documents of anyone at a customer, it offers no download of your data, and it has no way to sign in as you or as anyone in your company. The one exception is Peoplume’s own staff accounts: an owner-only list shows each one’s email, role, two-step status and the company it sits in. An owner also sees the note recorded with a deletion, which may quote the request.
A staff account opens the panel only with two-step verification on and a sign-in with its code within the last 12 hours. Support staff can only look. The panel’s one change — extending a company’s paid period — is an owner’s, recorded in that company’s own audit log with the staff member’s name. Nothing else is in the panel. Clearing two-step verification, removing a login nobody ever used, deleting a company on request, giving or removing Peoplume staff access, and emailing every company’s Administrators a notice we owe them (a new sub-processor, a breach) from a fixed text are each a command an owner runs outside the product, with a single-use code from their authenticator (except an owner’s access given with direct access to the database while no other owner can act). To send a notice, that command reads Administrators’ addresses inside the service and never prints them. Clearing two-step verification and removing a login are recorded in that company’s audit log under Peoplume’s name, and the company is told; a notice is recorded in the log of each company it reached; giving or removing staff access is recorded in the staff account’s own company and its holder is told; deleting a company is carried out as our privacy notice describes: a record of it is kept, and the backups holding it rotate out afterwards.
Those are the panel’s limits, not the limit of what we can reach: the people who operate the service can also reach the database itself — to back it up, restore it or fix a fault — and do so only where they need to (DPA §3).
Money moves need two people
Nobody approves their own leave, expense claim, travel advance, timesheet or pay proposal — enforced on the server in both directions, not by removing a button; nor hands over or settles their own travel advance; whoever files a claim on someone’s behalf never approves it (they may still reject it or send it back — stopping a payment stays one person’s call), and whoever requests an advance on someone’s behalf never approves or declines it (once someone else has approved it, they may still cancel it before it is handed over — stopping a payment stays one person’s call). Writing off a travel advance’s balance, or moving it to a loan, also needs permission to change salaries. Expense claims above a threshold you set need a second approver who is not the first. An offer needs a signature from someone other than the person who drafted it — unless the company had only one console account with active access when the offer was sent for approval, since nobody else could then sign it. Approved payroll runs are frozen snapshots, so editing a salary changes next month’s pay and not last month’s.
A company can require that opening a job, or changing its salary range, positions, branch or department, is approved by someone other than the person who asked (a company setting) — unless, when it was sent, nobody else could approve it. Each request is decided once and recorded; nobody decides a request to replace themselves. Completing an internal move — an employee hired into another job through an offer — needs recruitment, employee-editing and salary permissions and is refused on one’s own record.
One exception is deliberate. An hourly employee is paid only for a month that has been signed off, and when they have not submitted it, someone in HR can sign it off on their behalf — never on their own record — with a reason, and only for the hours they were shown: if the month’s attendance changes after they looked, the sign-off is refused and they look again. While that sign-off stands the month is marked as signed off by HR, and the sign-off is audited. That is one person’s signature, not two. The second look comes at payroll: whoever approves the run is shown each such month (who signed it off, when and why) and must confirm it before the run is approved. In a small company that may be the same person; the run then says “signed off by you” and its audit entry records it, which is a clear record rather than a second pair of eyes.
Files, integrations and outbound requests
- Uploads are limited by type and size, path-traversal guarded, and served only through routes that re-check who is asking — portal users can reach their own documents and nobody else’s, except a candidate’s CV, below.
- A candidate’s CV is downloaded only by users with recruitment access, by a member of the interview panel of a scheduled or held interview of that candidate, or by the opening’s hiring team in the employee portal — the last two never for their own application; an interview’s calendar file only by users with recruitment access or a member of its panel. Both are served private and uncached.
- The careers pages, the job feed and the careers logo are reachable only through the employer’s careers link. The feed and the logo are rate-limited per address before the link is looked up and per company after; an unknown link and a company that does not publish get the same answer from the feed. Careers pages ask search engines not to index them unless the employer turns on job-board publishing, which is refused during the trial. The logo is served only if its bytes are a PNG, JPEG or WEBP image.
- Moving several candidates at once applies each candidate’s own checks one by one, sends candidates nothing, and is recorded once in the audit log. Removing an application, or deleting a candidate, tells each interviewer of its upcoming interviews that it is off, never who the candidate was; anyone who cannot be told in Peoplume is named, with the time, to the person who removed it. The audit log records that a recruiter emailed a candidate, never the subject or the text.
- Webhook URLs, like every address we fetch from a company’s identity provider, must be HTTPS on a public host: local names and literal loopback, private, link-local, unique-local and IPv4-mapped addresses are refused; redirects are not followed, and the address is re-checked at send time rather than only when saved. This is to stop a webhook being used to reach inside our own network. A public name whose DNS points at a private address is not caught.
- API keys carry explicit scopes checked on the server, and a key can carry only data the person creating it may see themselves (salaries need permission to view salaries). Requests are rate limited per key, and per network address before the key is even checked. A key can be revoked, and a webhook paused or removed, at once — also while the company is read-only; a webhook can be subscribed only to events about data the person setting it up could see themselves when they set it up, and nobody whose role is limited to one branch can set one up. Each webhook delivery is signed with the endpoint’s own secret and carries its own delivery id; it is sent once and never retried, and events that happen while a webhook is paused are dropped, not sent later. Like webhooks, keys belong to the company rather than to whoever created them, so they outlive that person’s console access or a change to their role (see above).
- Push notifications, where a person turns them on for a device, carry the notification’s title and text, as the in-app notification shows them (which can name a colleague, a date or an amount), and an in-app link, and are encrypted to that device, so the browser vendor’s push service relays bytes it cannot read. A device’s subscription is a credential: it is stored for its own account only, never shown or exported, and deleted the moment the push service reports it revoked. Signing out of a browser turns that browser’s notifications off first, and a browser still holding another account’s notifications drops them when someone else signs in there.
- A praise notification carries no text and no author’s name — only that praise was received — and is deleted with the praise when it is retracted, removed or erased, or when the author’s or the recipient’s record is deleted, so no notification keeps a copy of the words. When an administrator removes someone else’s praise, the removal is recorded in the append-only audit log with the praise’s author and its opening words, and that entry stays.
- Notifications to the phone app travel through Expo and then Apple or Google, and are not encrypted to the device, so they say only which area changed — “Leave update”, “A payslip is ready” — never a name, a date or a figure. Each carries an in-app link, which names the page only where the page says nothing sensitive: for a one-to-one, an improvement plan or a probation the link is the app’s own inbox, so Expo, Apple and Google cannot tell from it that someone is on a plan. The app fetches the detail from us once it is opened.
What the browser is told
Every response carries a Content-Security-Policy, HTTP Strict Transport Security for a year, nosniff, X-Frame-Options: DENY, a referrer policy and a permissions policy. The permissions policy switches off camera, microphone, USB and the motion sensors outright, switches payment off on every page except the payment page, /billing/pay, where Paddle’s checkout frame may use it, and allows location only for our own pages — the portal’s check-in button asks for it, and nothing else does. (The phone app is not a web page: it asks the phone itself for the camera, only when an employee photographs a receipt.)
One honest limitation in that policy. Scripts are restricted to our own origin and Cloudflare’s analytics script (see the privacy page), except that the payment page also loads Paddle.js from cdn.paddle.com and frames Paddle’s checkout; so an injected <script src="…"> pointing somewhere else will not load, and the directives that usually turn an injection into stolen data — where forms may post (our own pages and the payment providers’ checkout and billing pages), what may be embedded, what a base tag may rewrite — are locked down. But inline script is still permitted, because the framework puts its own inline script in every page. Removing that needs a per-request token and a change to how pages are rendered, which is a deliberate piece of work rather than a setting, and we have not done it yet.
Backups
Backups are taken with an integrity check, verified by checksum, and copied off-site to object storage. Restores are scripted and default to a dry run, so recovering does not begin by overwriting anything.
The nightly backup runs on GitHub’s hosted runners, so each night’s copy of the database passes through GitHub, which also keeps it for 7 days as a build artifact. We do not encrypt that copy before it is stored either, so GitHub could technically read it too. Each day a second job on those runners reads the newest off-site copy back, checks that it opens whole and that every database migration it records is one the code has, and deletes it; that job keeps no copy.
They are encrypted at rest by the storage provider, not by us before they leave. The difference matters: it means Cloudflare could technically read a backup, where client-side encryption would mean it could not. We would rather state which of the two this is than let the stronger one be assumed.
What Peoplume does not do yet
Stated plainly, because you would find out eventually and it is better that you find out now:
- No SAML. Single sign-on is supported over OpenID Connect, which every major identity provider speaks — but if your procurement requires SAML specifically, we do not have it. That is a deliberate choice rather than a gap we have not reached: SAML security rests on XML signature verification, whose failure mode is accepting an assertion it should have rejected, and we would rather say no than implement it badly.
- No SCIM. Turning someone off in your identity provider does not end their Peoplume account; with Require single sign-on on they cannot sign in again with a password, but a session already open lasts until it expires (in a browser, 7 days at most; in the phone app, after 60 days unused) — sign them out of every device from Settings to stop it sooner.
- No independent security certification. Peoplume has not been through a SOC 2 or ISO 27001 audit and has not had an external penetration test. If your procurement process requires either, we do not have it.
- Almost no automatic deletion. Three things are deleted on a schedule: a leaver’s portal and app login the day after their last day (a console account is suspended on its own end date instead, never deleted automatically), sign-ups never confirmed once their link expires, and — only if the employer turns it on — rejected job applicants after the period they set, with the emails written to them and the notifications that link to their application (one who asked to be kept for future roles is kept until that request runs out). A former employee’s personal data can be erased on request from the day after their last day (on their own local day, not while an expense claim or air-ticket cash-out of theirs is still pending or approved but not yet paid, not while a travel advance of theirs is still requested, approved, handed over or being settled, and not while a console account tied to the record still has access — that is ended in Settings first): their identity, contact and bank details, documents, family members, the application they were hired from (CV and the emails written to them included) — or, tied to no record, a hired application under their email, when no other record not yet erased carries that email — and any on which they were an internal candidate, with the notifications that link to them, their places on hiring teams and their unsubmitted interview feedback drafts, receipts of claims never paid (rejected, including an approval later withdrawn, or sent back and not resubmitted), claims in trip reports they never sent (with their receipts) and those reports, with any declined or cancelled travel advance on them, the purpose and notes typed on their trip reports and travel advances, their name on checklist steps they ticked and on reviews they submitted about colleagues from a login tied to their record (released, or still held for calibration), and the reason typed on a stated leave balance or a pay hold, and the notes approvers and HR wrote on their leave requests go, as do praise written by or about them (with its notifications), their places as a successor, and the notes and flight-risk judgement on the succession plan covering them, their place as the head of a department, the certificate files on their training records, the note on each probation decision about them and the creator’s name on their own goals, their one-to-ones (shared and private notes and action items), suggestions of peer reviewers about or naming them, the note on a manager’s recommendation on their probation, and their name on what they recorded about colleagues as a manager, while payslips, attendance, leave (with their gender, where a leave type is limited by gender: it decides which leave they had) and its requests with their dates, days and reasons, and the receipts of reimbursed claims stay as the accounting record, and their review, self-assessment, 360° feedback, goal and training records stay, text included, as evidence in employment disputes — with their comment on a review, their self-ratings and answers, the reasons they gave asking for training, their probation decisions, and improvement plans about them with their check-ins. The names of staff recorded on records those staff acted on stay, like the names in the audit log. Everything else stays until someone deletes it. See Privacy.
- The public API is read-only. Write access would need the permission and branch-scoping model re-expressed for callers without a session, which is its own piece of work rather than a flag we have chosen not to set.
Reporting a vulnerability
If you believe you have found a security problem, write to [email protected] before telling anyone else, and give us a reasonable window to fix it. We would rather hear about it awkwardly than read about it later.
