7 · AuditEvery action, approval and session recorded
What the AI can and can’t do
The AI can
The AI can’t
Choose which tool to use next
Run anything you haven’t connected
Read results to explain them
See passwords, keys or tokens
Propose changes
Make a change without your approval
Plan and rehearse a dry run
Go off-plan after you approve it — any variance asks again
Use tools an admin has switched off, even if it asks for them
The AI’s output is treated as untrusted. Every tool call is checked against its schema, and names such as hosts, users and services are checked against strict patterns, so a reply from the model can’t smuggle in shell, PowerShell or LDAP commands.
Prompt injection
Content from outside — emails, web pages, log lines — could contain text written to trick an AI. Opsaido limits the damage:
Protection
Effect
Outside content is marked as data
Emails and similar content reach the AI labelled as untrusted, with the instruction never to act on what they say
Approvals still apply
Even if the AI were fooled, a change still needs a person to approve it
Dangerous actions are separate
Deleting, powering off or restoring needs a permission most people don’t have
Everything is audited
You can see exactly what was asked, proposed, approved and run
Logins and secrets
What
How it’s protected
Connector logins, API keys, bot tokens
Encrypted at rest with a key derived from your SECRET_KEY; decrypted only for the moment a connection is made
Sent to the browser?
Never. Forms show “stored — leave blank to keep”
Sent to the AI?
Never. The AI sees results, not how Opsaido signed in
256-bit random tokens stored only as hashes; HttpOnly, SameSite, Secure cookies
Two-step codes
TOTP with single-use recovery codes and replay protection
Backups
Encrypted with a passphrase you choose
Keep SECRET_KEY safe and backed up. Without it, stored logins can’t be decrypted, and anyone who has it plus a copy of the database could read them.
Connections
Protection
Effect
SSH host keys pinned
The key is saved on the first test. A changed key is refused, which stops a machine pretending to be your server
TLS verified
Certificates are checked. Add your own CA in certs/ca/ rather than turning checks off
Least privilege
Each connector uses its own account with only the rights you give it; read-only is enough to monitor
Gateways
A jump host needs no admin rights and holds no logins for the targets
The platform
Layer
Hardening
Container
Runs as non-root with a read-only filesystem, all Linux capabilities dropped and no-new-privileges
Database
On an internal network only; not reachable from outside
Remote desktop service
Reachable only by Opsaido itself
Browser
Strict Content-Security-Policy, X-Frame-Options: DENY, nosniff, HSTS over HTTPS
Requests
Changes need a custom header (CSRF protection); live connections check their origin
What leaves your network
Goes to
What
When
Your AI provider
Your question and the results of the steps
Each request. Use a local model to keep this on your network
The licence server
Licence ID, installation ID, version and counts only (number of machines)
A periodic check-in. Offline licences never check in
Services you set up
Alerts and replies to your notification channels (email, Telegram, Slack…); requests to cloud services you connect (Microsoft 365, Cloudflare…)
Only when you add them
Nothing else
No telemetry, no analytics, no host names, no logins
—
For the strictest setups, run a local model (Ollama, vLLM, LM Studio) with an offline licence, and only on-network notification channels. Then nothing leaves your network at all.
Your part
Do
Why
Turn on two-step verification for every admin
A stolen password alone isn’t enough
Give connectors least-privilege accounts
Limits what any mistake can touch
Keep dangerous-action approval for a few people
Deleting and powering off stay rare and deliberate
Put Opsaido behind HTTPS with your own certificate
Protects sign-in and sessions on the network
Back upSECRET_KEY and the database
So you can restore — and so you know where copies are