Security & privacy

Your data stays where it lives.

Where data lives at rest, we'll find it. Myrikeep finds sensitive data with an agent that runs inside your infrastructure. Files are read and classified where they live, and only labels, counts and file locations come back to the console. AI assist is off unless you turn it on — and even then, only short, redacted snippets leave.

How it works

Classification runs in your agent
Detection runs inside the agent container you deploy, next to your files. The console queues scan jobs; your agent does the reading.
Only labels, counts and locations come back
A finding records what was detected, how confident the detector is, and where the file is — never the file's contents or the matched values.
Credentials stay on your agent host
Passwords and keys for your data sources live in a local secret file your team mounts read-only. The console stores only an alias — it rejects secrets in connection settings.
Read-only access
The agent container runs read-only with all Linux capabilities dropped, and connects to your sources with read-only permissions. Scanning never writes to your stores.

What we store — and what we never store

The console holds the results of a scan, not the data the scan read. Everything the agent sends back fits into the left column — nothing else.

What Myrikeep stores versus what it never stores
We store We never store
Labels, counts and confidence for each file File contents or samples
File locations, provider item links and mail metadata (sender, received date, masked subject, Message-ID and attachment names) Passwords and keys for your data sources
Risk scores, scan history and the audit log Sharing URLs from your storage
Exposure categories and counts Names or emails of people with access — exposure is reported as counts only

AI assist: off unless you switch it on

An administrator turns AI assist on per workspace — never by default, and never silently. It exists for one job: helping label files the local rules could not.

Only short, redacted snippets.
When assist is on, the agent sends one short snippet — length capped by your setting — only for files the local rules could not label, with every detected match masked before it leaves.
Never stored or logged by Myrikeep, never used by us for training.
Snippet text is not written to our database or logs — the audit trail records counts only. Myrikeep never uses your data to train models. A snippet is sent once to our AI provider to suggest a label.
Switching it off is enforced by the server.
Turning AI assist off takes effect immediately and server-side: the classification endpoint refuses content from every caller, including jobs that are already running. No agent can bypass it.

Who can see it — without sharing names

For sensitive files, the agent also inspects sharing permissions — read-only — and reports them back as one of a few normalized categories. You learn that a file is shared with "anyone with the link" or "3 outside recipients", not who those people are.

When a permission check is missing, unreadable or unsupported, the file is shown as unknown. Unknown is never shown as safe: it stays flagged for review until a later scan can check it.

Reported as categories and counts
The agent sends the category and the number of visible entries — never the names, email addresses or sharing URLs behind them.
Sharing categories
  • Public
  • Anyone with the link
  • External recipients
  • Company-wide
  • Broad group
  • Restricted (no broad access found)
  • Unknown — needs review

Test Drive: deleted right after the scan

Test Drive lets you try Myrikeep on sample files you upload. The results page lists file names and the labels found — never file contents — and is reachable only with your private link.

Files deleted right after the scan
As soon as the demo scan finishes, your uploaded files are deleted from the isolated demo workspace.
Email confirmation
We send you an email confirming the deletion, so you have a record of when your files were removed.
Results kept for 24 hours
Results stay available with your private link for 24 hours; after that, everything — file names, results, your email address — is wiped.

Platform security

The controls behind the console — each one implemented and tested in the product today.

Strict Content Security Policy
Every page ships a nonce-based CSP with no inline scripts or styles. The only third-party script is Cloudflare Turnstile bot protection on sign-up forms, when it is enabled.
HTTPS with HSTS
Production serves HTTPS with HSTS and secure cookies.
Workspace isolation, tested on every page
Every workspace route requires membership in your workspace. An automated test walks the full route table to prove a forged workspace selector leaks nothing across workspaces.
Role-based access
Owners and admins manage connections, agents and settings; members can read the workspace but not change it.
Audit log of admin actions
Connections, agents, settings and other administrative changes are recorded in an append-only, per-workspace audit trail.
Agent credentials you can rotate and revoke
Rotate or revoke any agent from the console. Revocation immediately invalidates its credential and job tokens and requeues its running scans.
Per-job tokens
Each scan lease gets its own token, stored only as a hash and valid for a short, renewable window — so one job can never speak for another.

Frequently asked questions

No. Reading and classification happen inside the agent on your infrastructure. What comes back is what the scan found — labels, counts and file locations — never file contents.

Last updated October 16, 2026