Built to be checked.
Security and governance
The owner wants to know their automations are safe. You want to know how. This page is the how. If something you need isn't here, ask and we'll answer in writing.
The short version
- One workspace per client.
Your files, memory, skills and logins are isolated from every other client's.
- Your accounts, not ours.
StackProbe connects to your systems with your own accounts and only the permissions you grant.
- Secrets never travel.
Passwords and keys are encrypted and unlocked only for the moment a job needs them. They never appear in chat, in files or on a laptop.
- Every action is checked first.
A rules engine decides, per system and per action, what the AI may do. Read-only skills can never write.
- A person approves what goes out.
In Assisted mode, nothing leaves your business without approval.
- Everything is recorded.
A tamper-evident ledger of every run, action and change, chained so edits and deletions can be detected.
- Backed up, encrypted, off-site.
Every few hours.
- Your data is yours.
Export everything at any time.
Isolation
One sealed workspace per client
Each client runs in its own isolated workspace with its own files, memory, skills and stored logins. Jobs for one client cannot read, write or even see another client's workspace. There is no shared memory and no shared skill store across clients.
Staff sign in with their existing work account (Microsoft 365 or Google Workspace). StackProbe doesn't create a second set of passwords for people to lose.
Access and credentials
Your accounts, your permissions
StackProbe connects to your systems using your own accounts and the permissions you choose to grant. Where a system supports it, we ask for the narrowest scope that does the job (for example, read mail but not delete it). You can see what each connection is allowed to do in the console, and revoke it from your side at any time.
Passwords, API keys and tokens are encrypted at rest. They're unlocked only for the moment a job needs them, then locked again. They are never shown to the AI as text, never written into chat, logs or files, and never stored on anyone's laptop.
Our team's access to your workspace is limited to operating the service, is logged in the same record you can see, and can be restricted further on request.
Rules engine
Every action is checked before it runs
Before the AI does anything in one of your systems, the action is checked against your rules. Rules are set per system and per action, in plain English, and enforced in code. For example:
- QuickBooks: may read invoices and create draft purchase orders; may never void or delete anything.
- Shopify: may create draft orders; may never cancel an order or issue a refund.
- Microsoft 365: may read mail and draft replies; may send only with approval; may never delete mail or change mailbox rules.
Three defaults apply to every client:
- Least access. Each skill declares which systems it needs. It gets those and nothing else.
- Read-only means read-only. A skill marked read-only cannot send, change or delete anything, in any system.
- Loosening is loud. Making any rule less strict takes an extra confirmation and raises an alert to you and to our team. Tightening a rule is immediate.
A kill switch stops every job for your company in one click. Jobs stay stopped until a person turns them back on.
Approvals
A person decides what goes out
Every client starts in Assisted mode. In Assisted mode the AI can read, think and draft, but anything that leaves your business, such as an email, an order change or a payment, waits for a named person to approve it. Approvals arrive in Microsoft Teams (Approve, Edit, Reject), as a draft in the mailbox, or in a daily digest, and are recorded with who approved what and when.
Autonomous mode lets a specific job act without approval, within limits written down in advance: which systems, which actions, how much it may spend. You turn it on per job, after the job has proven itself in Assisted mode, and you can turn it off at any time.
The record
A ledger you can check
Every run, every action the AI takes, every approval and every change made in another system is written to an audit ledger. Each entry is linked to the one before it by a cryptographic hash, a fingerprint computed from the entry's contents and the previous fingerprint. Change or remove an entry and every fingerprint after it stops matching, so tampering shows.
The ledger is searchable from the console ("who approved the email to this customer?", "what did the low-stock job do on Tuesday?") and included in your export.
Monitoring and response
Watched every five minutes
Health checks run every five minutes and deeper checks twice a day. Failed jobs, broken connections, spending near a cap and backup problems alert our team, who respond and fix them. You see the outcome in your morning digest and, for anything that affected your data or your systems, a direct message from us.
Backups and continuity
Encrypted, off-site, every few hours
Your workspace, including files, memory, skills, rules and the record, is backed up every few hours. Backups are encrypted and stored off-site, separate from the systems that run your jobs. We test restores.
Data handling
Where your data goes, and where it doesn't
- What we store
- The files, memory, skills, rules, logins (encrypted) and records in your workspace, plus the data your jobs read from your systems while they run.
- Who processes it
- To run AI work, StackProbe sends the relevant content to an AI model provider under a business agreement that does not allow your data to be used for training.
- What we never do
- Sell your data, use it to train models, share it with other clients, or move it to a laptop.
- Your rights
- Export everything at any time. When you leave, we delete your workspace and backups after a short holding period and confirm in writing.
What we don't claim
We'd rather be precise than impressive. As of today:
- We don't hold a SOC 2, ISO 27001 or similar certification.
- We don't publish an uptime percentage.
- We don't claim our systems can't fail. We claim that when they do, we know within minutes, and it's our job to fix it.
If any of these change, this page will say so.
Questions to ask any vendor (including us)
- Can one client's automation ever see another client's data?Isolation →
- Where are my passwords and API keys stored, and who can see them?Access and credentials →
- What stops the AI from doing something nobody approved?Rules engine and approvals →
- If something was deleted or changed, how would I know?The record →
- Who gets the alert when a job fails at 3am?Monitoring and response →
- How do I get all my data out?Data handling →
Questions
Do you need admin access to our Microsoft 365 or Google Workspace?
Only for the permissions a job actually needs, and you grant them from your side. You can review and revoke them at any time.
Can our IT provider review your setup?
Yes. We'll walk them through it and answer questions in writing.
Can we restrict which staff can approve actions?
Yes. Approvers are named per job, and approvals are recorded against the person who gave them.
What happens if StackProbe shuts down?
Your workspace is exportable at any time, and your automations are documented in it.
Do you have a privacy policy?
Yes: privacy policy. It covers what we collect on this website and in the service.
Ask us anything about security
Email info@stackprobe.io with "Security question" in the subject. You'll get a written answer from a person.