# Custom roles, link safety, and a billing audit

6 September 2026. Custom roles now work end to end, with the organization's own database deciding what a member may do. Links BuildBase hosts are checked for reputation and can be killed at click time. And a systematic pass over billing fixed six real issues.

## Added

- **Access control:** Custom roles work end to end. The organization's own database is now authoritative for what a member may do, so the role you assign is what the API enforces, rather than the broader membership tier the person signed in with. On a self-hosted install that database is yours, which is the only place the answer can live.
- **Access control:** An audit log for the organization itself, in organization settings: who was invited and at what role, who accepted or declined, who was removed, and when support impersonated the organization. Readable by owners and admins only.
- **Access control:** An append-only security trail inside each organization now records every authorization denial, the full API token lifecycle, workspace membership changes, invitation rejections and revocations, secret rotations, and admin-API actions.
- **Security:** Links BuildBase hosts on your behalf are checked before a click goes through. A destination on the trusted list redirects straight away; anything else gets a reputation check and an interstitial that navigates itself once the verdict comes back safe.
- **Security:** Kill switches for hosted links. A single destination, or the tracking code covering every link in one sent message, can be blocked and refused at click time. That includes mail already sitting in an inbox: the message cannot be recalled, but the redirect can be refused.
- **Security:** A malicious verdict blocks the destination and disables the message carrying it, rather than warning only the one reader who happened to click. Both are audited, and the operator alert carries a one-click reversal, because false positives are certain.

## Improved

- **Console:** The console decides "may I do this" in one place instead of four, evaluating in the same order the server does, so the buttons and the API agree. A control you lack permission for is disabled with the reason on hover rather than hidden, because a control that vanishes cannot tell you what to ask for.

## Fixed

- **Access control:** A Slack incoming-webhook URL is a bearer credential: anyone holding it can post into the organization's channel. Any authenticated member could read it, whatever their permissions. The route now carries the guard it was missing.
- **Access control:** Revoking a permission could keep working for up to 60 seconds on any instance that had already cached it, including the one that made the change, so an admin who removed a permission watched the old one keep working. A change now reaches every instance immediately.
- **Console:** Screens and controls that reached data they did not gate on now check the permission they actually need, so a role with read access no longer sees write controls that answer 401.
- **Access control:** The roles editor no longer offers permissions the API rejects, and the audit screen resolves who did what instead of showing raw identifiers and truncated rows.
- **Email:** Phone, SMS and in-message anchor links keep working in tracked mail. A scheme allowlist added alongside click-time blocking had turned tel:, sms: and relative links into a blocked page, including in mail that had already been delivered and previously worked.
- **Email:** Open and click tracking links are signed with their own key rather than the one used for sign-in sessions. Existing links keep working, because verification falls back to the old key, which matters when a tracked link stays valid for a year.
- **Short links:** Archiving or deleting a short link actually stops the redirect, and a destination is validated when it is saved, so a javascript: or data: URL can no longer be stored and later emitted in a redirect.
- **Authentication:** Cloud-hosted auth domains serve the full security header set. The auth UI is a static export, so the headers configured for the self-hosted container were never served by the cloud deployment, which is the surface where custom auth domains are attached.
- **Billing:** Webhook events for a failed payment, a payment needing extra verification, and a successful payment could be silently dropped depending on which Stripe API version sent them - so the dunning emails and the usage reset after payment did not always run.
- **Billing:** Canceling a subscription could report the pre-cancellation state back for a few seconds right after the cancel succeeded, so a page that refetches on success looked like nothing had happened.
- **Billing:** Hitting a seat or trial limit returned a generic server error instead of a clear response an app could detect and use to prompt an upgrade.
- **Billing:** A workspace on the free plan could get stuck there and never be able to upgrade to a paid one.
- **Billing:** Rapid concurrent checkouts could create more than one Stripe customer for the same workspace, and a subscription on the extra one could go unnoticed by our duplicate-subscription checks.
- **Billing:** Checkout did not always prefill the buyer's own account email, so a trial could be started under a different address than the one on file.

Source: https://www.buildbase.app/changelog/billing-audit-fixes
