Security & data
This is the operational view for admins: what WikiFix touches in your Confluence and Notion, what it keeps, and what it doesn’t. The formal claims — jurisdiction, encryption, subprocessors, what we don’t have yet — live on the security page; this page summarizes and links rather than restates.
What WikiFix reads
Section titled “What WikiFix reads”Confluence: pages in the spaces a scan scopes, fetched read-only through Atlassian’s Forge platform at scan time — or, for a Confluence site added on the Sources tab, through Atlassian’s API with each person’s own Atlassian sign-in. Notion: the Notion pages a scan scopes, read through Notion’s API.
To show you what you can choose, WikiFix does list the names of the spaces you can see and, for Notion, the names and structure of what is shared with it; that stays in memory for minutes and isn’t stored, apart from the id of each space or page listed. Three things bound what a scan can reach: what it was scoped to (a scan never fetches the content of anything outside it); what the person who owns the scan can open; and what the app may read — in Confluence, the access your admin grants WikiFix at install, which they can restrict further with Atlassian’s app access controls; in Notion, only the pages the admin who connected the workspace shared with WikiFix. The Confluence scope list is public in the Forge manifest on the Marketplace listing; the Notion permissions are on Confluence and Notion.
What’s stored vs. processed transiently
Section titled “What’s stored vs. processed transiently”Processed transiently: the full text of scanned pages. Pages are fetched and analyzed during a scan run; WikiFix keeps no full copy of your wiki.
Stored: the working set WikiFix needs to do its job —
- text derived from your pages, which can repeat their wording and is kept whether or not it leads to a finding, and a numeric index of them,
- questions and answers on a finding, and comments you ask WikiFix to post,
- a search index, so you can search your findings,
- when a fix is previewed, the passage it would replace and its proposed text, so it can be applied and reverted — for a duplicate-pages merge, the whole text of each page merged away,
- scan and run history — kept in WikiFix’s own database and shown in the app, not written into your Confluence or Notion — and the credit ledger,
- for each person who uses WikiFix for Confluence, a sign-in record in Amazon Cognito named with their Atlassian account id, with no email address or password, and, while they sign in, that id and the site id for about five minutes in a DynamoDB table,
- for each person who signs in to a Notion workspace or an added Confluence site: their sign-in tokens, encrypted, and their account id there — plus, for Notion, the workspace’s id and name. WikiFix reads no email address from Notion.
All of it is stored in AWS Frankfurt (eu-central-1), and that is where the processing happens too, with the one exception below. The text in it is encrypted column by column. Your content is never used to train AI models, or to tune or improve our AI prompts. A person at WikiFix reads it only when you ask us to in writing or give explicit written permission, and only for the problem you raised. See the security page for the contractual detail.
The AI steps run on AWS Bedrock in the EU, apart from two analysis checks. With EU residency required on the Residency tab in WikiFix Admin, those two run on different AI models, which reside in the EU: eu/jev, operated by Bevelites GmbH on its own server in Germany, so every AI step runs in the EU; see Data residency. Without it, they run on TypeSafe, whose endpoint processes in the US, on every Data handling level. Either way, a request receives excerpts of the pages it checks, which can come from several pages: each passage with its page’s title and headings and the section around it, up to 2,500 characters of section. On a short page that can be the whole page; it never receives the whole space. eu/jev doesn’t store the excerpts it receives, in its database or its logs, and doesn’t use them to train AI models. We use TypeSafe with zero data retention: it doesn’t keep the excerpts it receives or use them for anything else, including training AI models. The other path that leaves the EU is yours to choose: on a paid Advanced plan you can add your own Anthropic API key, and the AI work then goes to Anthropic’s API under your account and your agreement with Anthropic, whatever residency you’ve set. Indexing stays on AWS Bedrock. Bring-your-own-key isn’t available on Standard, or during Atlassian’s 30-day evaluation; see Editions & plans and Credits & capacity.
One more thing worth knowing: on the Standard level, the text sent to the model and its reply are recorded in Langfuse, our monitoring sub-processor, so we can look into a problem when you ask us to. That text includes the page content under review, so unlike the analytics below it is genuinely content-bearing. Langfuse is pinned to an EU instance, and we delete that text about 30 days after it’s recorded. The full list of who processes what is on the sub-processors page, and changes to it are published in advance to a feed you can follow.
Data handling
Section titled “Data handling”Admins choose a level under WikiFix Admin → Data handling: Standard, Restricted or Strict. During onboarding, the first admin to answer is asked about sensitive data and EU residency, and later admins aren’t asked again. The team starts on Standard. WikiFix keeps what it needs to work on every level; the level decides what else is recorded, and the security page has the full table.
- Restricted records no AI text, and sends no site address, email, name or Atlassian or Notion account id to Sentry or PostHog.
- Strict also turns off Sentry, PostHog usage events and session replay. Feedback you choose to send still reaches us.
If your content holds special-category or criminal-offence data, such as health data, you must use Restricted or Strict. WikiFix doesn’t look at your content, so it can’t tell. Neither costs more per scan, but we can’t look into past AI results if you report a problem. Moving to either deletes your team’s records at Langfuse, usually within about a day and longer for a team with many records. The tab also shows each change of your configuration, permissions and consent, and each share link created, revoked or opened. It doesn’t list reads of your content.
What WikiFix writes
Section titled “What WikiFix writes”Nothing, until you click:
- Apply writes the chosen fix to the disagreeing pages — as the approving user, visible in page history, and revertable in one click. For a departed-owner finding, Apply reassigns the page’s owner instead of editing its text; that one isn’t revertable, because handing the page back to someone who has left doesn’t mean anything. The card tells you so before you click.
- Ask owner posts a comment on the page, anchored to the passage in question — on a Notion page, a comment on the page as a whole — and hands the decision over: the finding moves to your Escalated tab until someone settles it.
In Notion, both go through the clicking person’s own Notion sign-in.
WikiFix creates no pages in your Confluence or Notion — scan results live in WikiFix, not in your wiki. Merging duplicate pages doesn’t delete or archive anything either: the pages you didn’t keep stay live, their content replaced by a short notice linking to the one you kept, so existing links to them keep working.
On uninstall
Section titled “On uninstall”Comments WikiFix posted stay in your Confluence and Notion — they’re ordinary comments, and you can delete them like any comment. Server-side, scans of your site stop at once and the app token we cached for it is deleted. WikiFix marks the installation removed and holds the rest of your data for 30 days, so a re-install picks up exactly where you left off. Notion workspaces and other Confluence sites you added by hand stay connected for those 30 days, with their scheduled scans waiting; their sign-in tokens are deleted when the 30 days are up, and Notion is asked to revoke its own. After 30 days a daily job deletes your data from our database, whether or not you ask, once every install on your team is uninstalled; backups roll off 7 days later. Text in Langfuse is deleted about 30 days after it’s recorded, server logs expire after 90 days, and each entry of your change history, with the ids that tie it together, stays for three years from its date, even after your team is deleted, unless you ask security@wikifix.ai to delete it sooner. One thing stays that we do not control: the pending summary the WikiFix agent in Rovo may hold for a person in Atlassian’s Forge storage, which has no expiry from us while WikiFix is installed; Atlassian deletes app storage after an uninstall, on its own retention schedule. Want your data gone sooner? Email security@wikifix.ai.
Disconnecting a source on its own deletes none of its stored data: its findings and reports stay until the team is deleted.
The WikiFix agent in Rovo
Section titled “The WikiFix agent in Rovo”When someone uses the WikiFix agent in Atlassian’s Rovo, Atlassian’s Rovo AI receives the finding text, page titles and the passages behind them, the text of any page the agent reads for them and, when they settle a finding, the before and after text. Your pages are already at Atlassian. The Data handling level doesn’t change this: it is the same on Strict. To settle a finding the agent keeps one summary per person in Atlassian’s Forge storage until they settle it or prepare another.
What we collect about how you use WikiFix
Section titled “What we collect about how you use WikiFix”WikiFix sends product-analytics events to PostHog and error reports to Sentry, both stored in the EU, so we can see which parts of the product work and which break. Events from our servers pass an allow-list in code: a property that isn’t on the list for that event is dropped. Events from your browser pass a block-list: page titles, excerpts, space names, and Atlassian account IDs are blocked outright. Events carry counts, durations, outcomes, our own internal IDs, page addresses, browser and approximate country. Sentry reports carry stack traces and error codes, with no page content, names or emails, and Sentry also gets performance traces: the timings of requests and page loads, with addresses but no query strings.
Session replay masks text and inputs, leaves out elements that hold your text in an attribute, and records layout and clicks.
One thing to know: comments you choose to write in feedback reach us exactly as you wrote them, and on the Standard level so does your reply-to address if you leave it ticked, so we can read them and answer you.
To turn these down, see Data handling.
The formal version
Section titled “The formal version”Encryption in transit and at rest, tenant isolation, subprocessor list, what we don’t have yet (and when we will) — all on the security page. The short privacy policy covers the same ground from the data-subject side. Our Data Processing Agreement applies when you accept the Terms, so there’s nothing to sign. Anything unclear: security@wikifix.ai.