The short version
Inbox Imp holds access to your mailbox. That is a serious thing to be trusted with, and this page describes what we actually do about it, including the parts we have not done yet. Everything below is a description of the running system, not an aspiration.
Access to your mailbox
- OAuth only. We never see or store your password. Access is granted through Google's or Microsoft's own consent screen and you can revoke it at any time from that account.
- Verified by Google. Our use of the restricted
gmail.modifyscope passed Google's OAuth verification, including a CASA Tier 2 security assessment. - Outlook and Microsoft 365. Microsoft mailboxes connect through the Graph
API under
Mail.ReadWrite,Mail.SendandUser.Read. Personal Microsoft accounts and work or school (Microsoft 365) accounts can both connect. - Tokens encrypted at rest. Refresh tokens are encrypted with AES-128 (Fernet) before they reach the database. The key lives in Google Secret Manager, separately from the data. Only ciphertext is stored; the plaintext token is never written down.
Your mail content
- Full bodies are never stored. We keep roughly the first 2,000 characters as a search preview, which for a short message is the whole message, so we say so rather than implying otherwise.
- Never used for training. Every AI request carries a no-retention instruction that restricts routing to providers who do not keep or train on your content. This is enforced in code on every path that reaches a model, not stated as a policy and hoped for.
- Stored in the United States. Application services in Google Cloud
us-central1, database and search index inus-east-2. AI inference is not region-guaranteed, and we do not pretend otherwise. - Attachments are read, not kept. Text and images from attachments are sent to the model so a message can be classified correctly. We keep a short description of each attachment with the decision; the file itself is never stored.
- Sender research touches the public web. "Research this sender" sends the sender address, display name, domain and subject line (never the body) through OpenRouter to a web-grounded model.
- Encrypted in transit with TLS everywhere.
How we avoid losing your mail
The failure that matters most in a product like this is not a breach. It is quietly archiving or deleting something you needed. We treat that as a security property:
- We never erase mail ourselves. Archiving is the default, and archived mail stays searchable in your account forever. Mail the assistant is confident is spam is moved to your provider's Trash / Deleted Items folder. It is recoverable there for 30 days, after which your mail provider empties it. Everything else stays where you can find it.
- Deterministic guards, not prompt instructions. Login codes, replies from real people, financial mail, security notices and future commitments are protected by code that runs after the AI decides and can override it. Each guard came from a real incident, and each has its own off switch.
- Everything is undoable. Every action in the app has an undo, and bulk operations report partial failures instead of rounding them off to success.
Infrastructure
- Google Cloud Run, with the job-running tier private (internal ingress, authenticated calls only)
- Row-level security enabled on every database table
- Payments handled entirely by Stripe and Apple; we never receive card details
- A daily automated tripwire on database traffic, so an unusual access pattern surfaces within a day
What we have not done
A security page that lists only strengths is marketing. These are real gaps, stated so you can weigh them:
- No independent penetration test has been performed. One is scoped and planned.
- No SOC 2 or ISO 27001 certification. We are a very small company; these are expensive and we have not done them.
- No stated maximum retention period for mail-derived data yet. Today it is kept while your account is active and deleted within 30 days of account deletion.
- No 24/7 on-call. Inbox Imp is operated by one person, so an incident at 3am is found in the morning.
- Microsoft publisher verification is complete; attestation is not. Bitwise Learning Inc. is a verified Microsoft publisher. Microsoft's Publisher Attestation and Microsoft 365 Certification have not been started yet.
Reporting a vulnerability
Email security@inboximp.com with enough detail to reproduce. We will acknowledge within 3 business days. Please give us a reasonable window to fix before disclosing publicly. We do not currently pay bounties, and we will credit you if you would like to be credited.
If you believe customer data is actively exposed, say so in the subject line so it is triaged first.