Home-lab security: do the few things that prevent most incidents
The realistic risk for a home AI server is exposure, weak passwords, stale images and missing backups. Work through exposure, auth, secrets, containers, segmentation, backups and incident response.
Threat model: you are an ordinary target
A home lab rarely faces a custom attack. The common cases are:
- a public scanner finds a service with no authentication;
- a container image is months old and carries a known issue;
- a password is reused, or an admin port is exposed;
- a disk fails, or data is deleted, before anyone checks a backup;
- a test environment keeps its debug endpoint open.
This does not mean advanced attacks do not exist. It means fixing cheap mistakes pays more than chasing a complex design.
Exposure: if a port is not needed, do not open it
- Prefer Tunnel or a VPN over inbound port forwarding.
- If a port is unavoidable, open only what is needed, restrict sources, isolate the service.
- Do not expose router, NAS, Proxmox or Docker admin interfaces.
- Make services listen on internal addresses; use a reverse proxy or tunnel for public access.
- Keep a list per public service: hostname, port, auth method, owner, last update.
Quick checks:
ss -tulpn | sort -k5
nmap -Pn -p- your-public-ip # run from another machine
An external scan is more reliable than “I think it is closed”.
Authentication and authorisation
- Enable MFA on admin services; disable password login where possible.
- Use a different password per service and a password manager.
- Create separate tokens for automation, with limited scope and expiry.
- Entrance auth and application auth are different: Access protects the door, not the app’s roles.
- Remove unused accounts and tokens on a schedule.
Secrets
- Never commit
.env, private keys or Cloudflare tokens. - Use the runtime’s secret mechanism rather than baking keys into an image.
- Inject via environment or mounted files with tight permissions.
- Split tokens by purpose: read-only, write-only, single resource.
- After a leak, rotate first, then investigate; deleting the commit is not enough.
Host and containers
- Pull from trusted sources; pin versions or digests instead of betting on
latest. - Run containers as non-root and drop unneeded capabilities.
- Mount only the directories that are needed; never mount
/or the Docker socket casually. - Set resource limits so one model cannot take down the host.
- Update base images and applications; read the notes before updating, verify after.
Minimum checklist:
- Image versions are pinned and you can explain why.
- Containers are not root, or have no extra privileges.
- Data directories are backed up and have been restored once.
- Logs do not contain tokens, passwords or full request bodies.
- Public ports have authentication; admin ports are not public.
Network segmentation
Split devices into at least three groups:
- Work/personal: laptops, phones.
- Lab/servers: models, containers, NAS.
- IoT/guests: cameras, TVs, visitor devices.
Use VLANs or separate SSIDs where possible; otherwise firewall rules that restrict cross-network access. IoT devices update slowly and should not reach server admin ports.
Backups: a backup that was never restored is a hope
- At least two copies, one offline or immutable.
- Restore one on a schedule; record duration and missing pieces.
- Cover config, databases, model files, certificates and keys.
- Never keep every copy on one disk, one machine or one account.
- Verify integrity so the backup itself is not corrupt.
A simple 3-2-1 variant: local copy, NAS copy, offline disk or object storage copy.
Logs and alerts
- Record failed logins, permission changes, config changes and backup results.
- Do not record passwords, tokens, identity documents or full chat content.
- Set retention so logs do not grow forever.
- Keep alerts few and precise: disk full, backup failed, certificate expiring, service down.
- Read the logs occasionally so alerts do not become noise.
Incident response: stop the bleeding first
When something looks wrong:
- Isolate: cut the network or close the entrance; preserve evidence.
- Scope: which accounts, tokens and data may be affected.
- Rotate: passwords, API tokens, SSH keys, cookie secrets.
- Restore: from a verified backup, and check that no backdoor came back.
- Review: timeline, root cause, improvements, owner.
Do not investigate while the service is still exposed, and do not restore without rotating credentials.
Pre-launch checklist
- Only required public entrances; admin ports are internal.
- MFA on admin accounts; passwords are unique.
- Image versions pinned; updates recorded.
- Secrets not in Git or image layers.
- Backups exist and have been restored.
- Logs exclude sensitive data; retention defined.
- A written “who to call, what to do first” flow exists.
Related notes
- The request path: Cloudflare basics.
- System and disk layout: system choices.
- Hardware sizing: run the assessment on the home page.