Security
What GrowthFabric's agents can and cannot do to your ad accounts, where your data sits, who can read it, and what we have not built yet.
Last updated 2026-10-21
This page describes the system as built, with the change code listed as of 21 October 2026. GrowthFabric is not open to customers yet. When we last checked our live server, on 14 October 2026, the model provider's key and error reporting were active; the email, payment and Meta keys and the key that signs approvals were not set, so changes to ad platforms are switched off. Where something is built but not switched on, we say so.
Where things stand
| Control | State today |
|---|---|
| Agents change your ad platforms | Off. Four separate locks hold it shut (below). |
| Read-only by default, per connection | In place |
| A named person approves each change, tied to that exact change | Built, inactive while changes are off |
| Platform tokens encrypted at rest | In place |
| Every database table protected by row-level security | In place |
| Customer app holds no privileged key | In place |
| Figures from incompatible sources refused, never added | In place |
| Deleting a workspace or connection | On request: you ask by email and we carry it out. Not offered in the app |
| Independent certification (SOC 2, ISO 27001) | Not built. None held |
| External penetration test | Not done |
| Single sign-on, multi-factor sign-in | Not built |
A filled diamond means in place today. An outline means off, not built or not done.
What agents can do to your accounts
Every connection is read-only unless a person sets it otherwise, and nothing in our code sets it. The setting is stored per connection and the database refuses a read-write setting without the name of whoever set it.
No agent can change anything on any platform. Four locks must all be opened, each on purpose:
- The key that signs approvals is not set on our live server (checked 14 October 2026). Without it no exact change is built, and every approved action comes back "not executed" (or "proposal only" when no exact change exists).
- A global write switch ships engaged. If our server cannot read the switch, or the row is missing, it counts as engaged. Releasing it takes a named person.
- No agent holds the trust level that may carry out a change, and no code grants it. Today we would have to write that grant by hand.
- No connection is set to read-write.
If the locks are opened later, two platforms have code to make a change, and only these kinds exist. On Google Ads: adding an exact-match negative keyword, pausing or enabling, pausing a Performance Max, Demand Gen or App campaign, and creating a Search campaign that starts paused. On Meta: pausing or enabling, shifting budget, and building a campaign that starts paused. All of it is off until you switch it on for a connection, and every change then waits for your yes (or, in auto mode, must sit inside a mandate you set; see trust levels below). None of these requests has yet been exercised against a live ad account. Google Analytics 4, Search Console and AppsFlyer are read-only, and the code refuses any change to them. For Google Tag Manager we never publish: a fix would be prepared in a new workspace for you to review and publish yourself.
What a change must pass
When changes are on, a change runs only after all of these hold:
- A named person approves that exact change (or, in auto mode, the change sits inside a mandate a named person set). The approval is a signed token over a fingerprint of the change. A forged token, an edited change or a token used twice is refused. Only the agent's attached owner may approve, an agent cannot approve, and the approver must belong to the same workspace.
- The change comes from a finding with evidence. A change with no finding behind it is refused, as is a change that does nothing, a duplicate, or a new ad or campaign that is not created paused.
- The write switch is checked twice, once before anything is read and once just before sending.
- The account is read again and the value must match. If the value changed since you approved, nothing is sent. Approvals have no clock: a proposal is withdrawn when the finding behind it stops holding, and the fresh read catches anything else.
- The platform's own dry run passes before the real request goes out.
- Your limits hold. The per-day change count and budget bounds are set per workspace, and an unset bound refuses. Today we set them by hand with you; there is no self-service setting yet.
- After sending, the result is read back. If it does not equal what you approved, it is recorded as an incident, a reversal plan is made, and that kind of change drops to a lower trust level.
No approval can override a short list of changes the code refuses outright: billing and payments, user access and credentials, removing or deleting anything, changing live tracking, uploading or removing conversions, changing attribution, how conversions are counted or their values, sharing conversion data, and accepting a platform's terms.
The record of changes
Every step of a change (approval, each check, the request, the response, the read-back) goes to a change ledger that no application role, including our own server, can edit or delete. The one exception is the anonymising step when a workspace is deleted, and a database owner could still drop the guard itself. Workspace members can read it. While changes are off, it holds no rows.
Trust levels
Each agent has a trust level, set per kind of change, from 1 to 5. Today we set them by hand with you; there is no self-service setting yet. Levels 1 to 3 never change a platform. At level 4, each change waits for your yes. Level 5 is auto mode: the agent acts only inside a mandate you set, which is versioned and has limits, and an agent can never create or widen its own mandate. A change outside the mandate needs a fresh yes or is refused. No agent holds level 4 or 5 today.
| Level | Name | What it adds |
|---|---|---|
| 1 | Observer | Reads and reports. |
| 2 | Assistant | Proposes changes and follow-ups, and waits for your approval. |
| 3 | Operator | Also runs checks, asks other agents for help and writes reports. No external change. |
| 4 | Trusted operator | May carry out a change only after you approve that exact change. Each change waits for your yes. |
| 5 | Autonomous | Auto mode: acts only inside a mandate you set (versioned, with limits). It cannot create or widen its own mandate. |
Promotion past level 3 is a person's decision. A write incident drops an agent's level, and nothing in the code restores it.
How well the write gate is tested
The gate has property tests (no generated sequence of events writes without a matching approval; a forged or edited token never writes; a token used many times at once writes once) and a mutation-testing run that fails if any change to the code survives the tests. That run is started by hand and is not part of our automated build; the last recorded result, on 30 September 2026, was every change caught. We have not run a live change.
How you connect your accounts
You sign in at Google or Meta on their own consent screen. We never see a platform password and never ask you to paste a token or key.
| Platform | Permission we ask for | What it allows |
|---|---|---|
| Google Ads | adwords | Google offers no read-only version for this API. It also allows changes at Google's side; ours are switched off as above, and our queries are read-only by construction. |
| Google Analytics 4 | analytics.readonly | Reading only. |
| Google Tag Manager | tagmanager.readonly | Reading only. |
| Search Console | webmasters.readonly | Reading only. |
| Meta | ads_read | Reading ads only. Connecting Meta is not open yet. |
The Google connection also asks for openid, to identify the person who signs in, and Meta may take its permission from a login configuration we set up in Meta.
You can withdraw access at any time in your Google Account. When you disconnect an account in GrowthFabric, our server asks Google to withdraw the token, then deletes the connection and its data even if Google's answer fails, and then tells you to check your Google Account. Meta is different: we delete our copy of the token, but nothing yet tells Meta to withdraw it, so remove GrowthFabric in Meta's business integrations too.
Signing in to GrowthFabric
You sign in to GrowthFabric with an email link or an email and password, through Supabase Auth. The session lives in a cookie that scripts cannot read, is sent only over HTTPS, and does not travel on requests from other sites. There is no single sign-on and no second factor yet.
Tokens and secrets
- Long-lived platform tokens are encrypted with AES-256-GCM, a fresh random value for every encryption, before they reach the database. Each row records which key version sealed it, and our server re-encrypts older rows whenever a key is replaced, so an old key can be retired.
- For Google we keep only the long-lived refresh token. Short-lived access tokens are requested for each call and never stored.
- The keys sit in our server's secret store, apart from the database. The table holding tokens has no access rule for any signed-in user, so only our server's own credentials can read it. Our server decrypts a token in memory for the moment a read needs it.
- The audit trail records that a token was stored, replaced or removed, never the token itself.
- The code cannot stop someone who holds both the database's server key and the vault key from decrypting. They are separate secrets in separate accounts.
- Sign-in tokens for a customer's AI assistant (see below) are random values held in Cloudflare's key-value store with short lifetimes: an hour for access, up to 30 days for refresh, rotated on each use, with the whole chain revoked if an old one is replayed. They are not hashed at rest.
Keeping workspaces apart
- Every table in the database has row-level security switched on, and a test on the real migrations checks that one workspace's owner reads nothing of another's. A person in no workspace reads nothing.
- Five tables have no access rule at all, so no one can read them from a session, only our server: platform tokens, two tables that support the write gate, the payment-event log and the scheduler heartbeat.
- Members read their workspace's data. Writing is narrower: most tables change only through our server. A few settings (alert thresholds, rule parameters, report boards, a finding's status) can be changed by an owner, admin or operator, brands by an owner or admin, and a viewer can read and give feedback on an output but cannot change settings, findings or boards.
- The website and the customer app hold no privileged key. Your browser talks only to the app; the app talks to the database under your own session and to our server with your own sign-in. Only our server talks to the ad platforms. We check this in tests: the app code never reads the database's server key.
- Our server takes the workspace from your verified sign-in, never from what an agent or a model says.
What the AI model sees
Our agents run on Claude, through Anthropic's commercial API in the United States.
- Detection is deterministic code, and the code computes every figure. The model writes the words around the figures, and its text is checked against them before you see it.
- The model sees structured, bounded, platform-reported data and your questions: findings and their evidence rows, search terms behind a waste finding, account names and ids, your Playbook, journal notes, open proposals, and for Meta a bounded set of campaign, ad set and ad tables (at most 25 rows and 31 days). It also sees what you type in chat.
- We do not strip names or other text from what we send, and there is no raw-query tool for our own agents.
- Anthropic's commercial terms say it may not train its models on customer content (Commercial Terms, section B). Anthropic's own documentation says it deletes API inputs and outputs within 30 days by default, longer in cases such as content flagged for policy violations. These are Anthropic's statements, not ours to guarantee. We have no zero-retention agreement.
- If you connect your own AI assistant to our server, that assistant calls our read-only tools and receives the rows directly. Which model vendor it is, and its terms, are your choice.
Numbers that do not mix
Two figures are added only by one function in our code, and it refuses when the figures are not comparable: different currencies, different platforms, different conversion definitions or attribution bases, a definition that changed inside the window, or overlapping dates. It names the reason. It never adds them with a warning. Property tests run against it with generated inputs.
What we keep, and where
The database is Supabase, in Frankfurt (eu-central-1). Our server runs on Cloudflare's network. The website is on Vercel's. The sub-processors are listed on the Sub-processors page.
| What | Where | Who can read it | Kept |
|---|---|---|---|
| Platform tokens | Database, encrypted | Our server only | Until you disconnect or delete |
| Account names, ids, structure, daily account-level totals | Database | Your workspace's members | While the account is connected |
| Findings, evidence rows, proposals, approvals, reports | Database | Your workspace's members | While the workspace exists |
| Chat messages | Database | Members, for threads under a brand of their workspace | While the workspace exists |
| Playbook, journal, settings, answers | Database | Your workspace's members | While the workspace exists |
| Live campaign, keyword and ad-level rows | Our server's memory for the request | The request | Dropped when the request ends |
| Billing records | Database, from the payment provider (not active yet) | Your workspace's members | With the workspace |
| Email addresses and sign-in | Supabase Auth | You, and our server | While the account exists |
| Audit and change records | Database, append-only | Members | Kept in anonymous form after deletion |
| Our server's logs | Cloudflare | Us | Cloudflare's log retention |
We read your campaign, keyword and ad-level figures when an agent needs them. We do not store raw platform responses, audiences, images or videos.
Logs and errors
- An append-only audit log records sensitive changes in the database (members, connections, token changes, trust levels, schedules, write limits, proposals), with secrets and message contents left out. Workspace members can read it.
- Our server's logs hold counts, ids and short error strings, not request bodies.
- Error text is passed through a scrubber that removes platform and sign-in tokens, email addresses, account ids, phone-like numbers, long digit runs and webhook addresses before it can leave our server. It cannot remove free text, such as a campaign name inside an error message. Scrubbed error reports go to PostHog in the European Union, from our server only, with no person profile.
- Each outage alert names the part of the service that is down and nothing else.
Deleting your data
Today we carry out deletion ourselves, on request, and confirm to you when it is done. For a workspace we ask Google to withdraw your tokens, delete the workspace and what is under it, and remove sign-ins left with no workspace. A copy of your data in one file is available beforehand to an owner or admin; a very large table is cut at 10,000 rows and the file says which.
What stays after deletion: audit and change records with the workspace, the people, the accounts and the contents removed from them (internal reference ids remain), and a record that the deletion happened, with the id of who asked. Our database provider's backups age out on its own schedule. Anthropic's copies follow its terms, above.
Deletion is not in the app yet, because the app is not open. Until it is, email dialog@growthfabric.ai from the address you use with GrowthFabric. The data deletion page has the steps.
Not built yet
We would rather say so:
- No SOC 2, ISO 27001 or other independent certification. Several of our providers publish their own; that does not make us certified.
- No external penetration test.
- No single sign-on, no multi-factor sign-in, no IP allow-list.
- No staff console with logged access. Only the founder can reach production systems today. The audit log records changes to sensitive tables in the database.
- No audit-log export. A workspace export in one file exists for owners and admins, without tokens and alert destinations.
- No hashing of AI-assistant sign-in tokens at rest.
- No withdrawal of the token at Meta when you disconnect Meta, and no matching of Meta's deletion callback to stored Meta connections. Meta is not open to customers.
- No change code for any platform beyond the kinds listed above, and none of it has run against a live ad account.
- No scrubbing of the text sent to the model, and no zero-retention agreement with the model provider.
- No security.txt file and no bug bounty.
- The status page is built and not published yet.
Report a problem, or send us your questionnaire
Email dialog@growthfabric.ai for a security problem. Please give us reasonable time to fix it before you share it publicly.
If your team has a security questionnaire, send it to the same address. We answer from this page and the data we hold, and where the honest answer is "not built", we say that.
See also the Sub-processors, the Data Processing Addendum, the Privacy Policy and the data deletion page.