Security & privacy
The basic principle is simple: Saka only manages. Your data stays on your server.
Saka does not store your data
Saka Panel does not store or relay user data. Website files, database contents, replication streams, backups, AI keys, and the assistant's chat history and memory all live on your server. Load balancer traffic does not pass through Saka either.
What Saka stores (metadata)
| Stored by Saka Panel | Not stored by Saka Panel |
|---|---|
| Account (email, password hash) | Database passwords |
| Server list, IP, OS, latest status, and the name of the network that owns the IP (e.g. VPS provider or ISP) for internal statistics | AI API keys, Telegram bot tokens, OpenClaw gateway tokens |
| Names and members of database clusters and load balancers | App log contents |
| Task history (command, a summary of inputs without secrets, result) | Database contents, files, backups |
| Alert history, Telegram notification connection | Metrics history (kept for 30 days on your server) |
| API tokens (stored encrypted, only the last 4 characters are shown) | WireGuard private keys |
Secret inputs used when installing an app only pass through to the agent. The history only records a summary, with secret values masked (•••).
Passwords are read from the server
- Databases: the password is generated when the cluster is created, passed once to the data servers, and then not stored. The Tampilkan sandi (Show password) button reads it directly from a data server at that moment. MCP never hands out passwords.
- AI keys: stored only on your server, in a file readable only by root. The model list is requested from the AI provider by the server, so the key never leaves it.
- Logs: read directly from the server on request, not recorded by Saka.
TLS & networking
- The agent connects outbound to Saka Panel over a secure WebSocket. No control or panel port is opened on the server.
- The agent is downloaded with a SHA-256 checksum check. When it updates itself, it only accepts releases signed with an Ed25519 key that is not stored on Saka's servers (the public key is built into the agent). Updates without a valid signature are rejected, so even a compromised Saka Panel cannot push a fake agent.
- Databases: TLS is required (PostgreSQL
hostsslwith SCRAM, MariaDBrequire_secure_transport). Traffic between locations goes over encrypted WireGuard. - Ready-to-install apps: ports only on
127.0.0.1. Dashboards open through an outbound tunnel with a one-time ticket, on a subdomain separate from the panel. - New ports are opened only when you create a database cluster (database ports and the private network between locations) or a load balancer (80/443 on the balancers, 53 for LB DNS), and only when the Saka firewall is on.
Firewall
The firewall is turned on only when you ask, after Saka shows you the ports currently in use. SSH is always left open. Afterwards, the agent makes sure apps can still reach the internet; if they cannot, the firewall is turned off again. On load balancer target servers, the app port is opened only for the balancer IPs. Details in The Readiness card. Once on, the Firewall card on the server page lists the open ports: Open port for a new app (the firewall stays on), Close for ports no longer used, and Turn off when needed (with a confirmation). SSH, Saka site, VPN, and platform ports are marked required and cannot be closed from the panel.
Forensic audit (stage 1)
The Forensic audit card on the server page snapshots the sensitive state of the server a few minutes after the agent starts: users and sudo rights, SSH keys per account, cron, systemd units, listening ports, SUID files, Docker containers and images, key config files (sshd, hosts, resolv.conf, passwd, shadow, ld.so.preload, rc.local, PAM), and the package list. Every hour the server is compared with that baseline. Changes made by Saka tasks are folded into the baseline silently; anything else (you over SSH, an AI agent, an intruder) is listed on the card and sent to Telegram, for example "SSH keys: added root: ssh-ed25519 …". Package changes are only recorded. Everything is read-only: nothing is modified while snapshotting. Check now compares immediately; Set as baseline is for after you reviewed the changes and are sure they were all yours. This is evidence for an internal investigation, not legal evidence: a valid chain of custody needs procedures beyond Saka's reach.
Event timeline (stage 2). On the same card, pick 24 hours, 3 days, or 7 days to see events in order: SSH logins per user and IP with failed-login counts, sudo commands, Docker container events, packages installed or upgraded, boots, system errors, changes outside Saka, and Saka tasks (who ran them, through which channel). Everything is read from logs already on the server; shell history is deliberately not read since it has no timestamps and anyone can edit it.
Evidence bundle (stage 3). Pick a range from 24 hours to 30 days and press Create evidence bundle: one zip with the baseline, changes, timeline, raw log copies (sshd, sudo, system errors, docker events, dpkg), and a laporan.md report readable by non-technical people. Every file inside has a SHA-256 (SHA256SUMS), that hash list is signed with the agent's Ed25519 key (the private key stays on the server and is never sent to Saka), and the public key plus the zip hash are recorded in Saka when the bundle is created. VERIFIKASI.md inside the zip lists the sha256sum and openssl pkeyutl commands to verify it. Deleting the file from the server keeps the hash record. The zip also contains laporan.pdf. For an auditor or board member without a Saka account, Share link creates a read-only link (valid 7 days, revocable) with the summary, hash, the public key as recorded by Saka, verification steps, and a download button; downloads through the link are logged in the action history.
Server security
The Server security card on the server page gives a score from 0 to 100 and a VPS checklist. Every row that is not safe yet has a Fix button:
- No SSH password login. Fix turns off SSH login with a password (including root), only when an SSH key is already installed, so you cannot lock yourself out. Connected sessions stay open, and the VPS provider's web console still works.
- Firewall on, with the ports you choose (same as on the Readiness card).
- No pending security updates, and daily automatic security updates (unattended-upgrades or dnf-automatic, no automatic reboot).
- No reboot pending after kernel updates.
- Saka Shield on.
AI agents read this list through server_perbaiki kind keamanan-cek.
Saka Shield
A server-level guard that does not depend on your app (WordPress, Laravel, Git apps, n8n, and SSH too). New servers: on automatically from install, with your IP at the time you created the install command (or pressed Create server in a cloud account) excluded. Existing servers: turn it on from the server page, Saka Shield card. The agent reads SSH and web access logs every 20 seconds and blocks IPs that within 10 minutes:
- fail SSH login 6 times (unknown user or wrong password),
- scan for weaknesses 6 times (e.g.
/.env,/.git, phpMyAdmin, phpunit), - submit a login form 30 times (
wp-login.php,/login,/users/sign_in, and similar), or - request 80 addresses that do not exist.
Blocks are applied in nftables for every port and Docker container, and in Caddy for visitors coming through a load balancer, tunnel, or Cloudflare (their real IP is read from trusted headers). Blocks last 60 minutes; IPs blocked again within a day are blocked for 24 hours. Blocks expire on their own even if the agent stops.
So you are never locked out: your IP when turning it on from the panel is excluded automatically, as are connected SSH sessions, private IPs, Cloudflare ranges, the server's own IPs, and load balancer nodes. Add your office IP under Exceptions; a wrong block can be lifted with one click. In a real test, an attacking VM was blocked on every port less than 45 seconds after 8 SSH attempts, and again after probing 8 weakness paths. MCP: perisai_status, perisai_atur, perisai_lepas.
Advanced settings
- Rate limit per IP: IPs that send more than 600 requests per minute (adjustable from 300 to 3000, or off) to your sites are blocked temporarily. Regular visitors stay well below that.
- Block countries: enter two-letter country codes (e.g. CN, RU). Their IP ranges (from ipdeny.com, refreshed weekly) are loaded into the server firewall; visitors through Cloudflare are blocked by Cloudflare's country header. By default only web ports, so you can still SSH while abroad; tick the SSH option to cover every port. Do not block the country search engines crawl from (e.g. US) if your site needs Google.
- OWASP WAF: the Coraza web application firewall with the OWASP Core Rule Set (SQL injection, XSS, path traversal, and more), for every site on the server. Start with Monitor: requests that would be blocked are listed on the Shield card (attack type, IP, address, rule) without blocking. If real visitors are not flagged for a few days, switch to Block: malicious requests get a 403, and repeat attackers are blocked in the firewall too. Adds about 50 MB of RAM (flat, regardless of the number of sites); the web server restarts for a few seconds when the WAF is turned on or off.
In real tests: 900 fast requests from one IP were blocked after passing 600 per minute; blocking Singapore closed the web for an attacking VM in Singapore while its SSH stayed reachable and visitors from Indonesia were unaffected; the WAF in block mode rejected SQL injection, XSS, and path traversal (403) without affecting regular pages and searches.
Database fencing
A database node that is not fit to accept writes closes its own database port to outside traffic (TCP RST), checked every 2 seconds. This stops apps from writing to the wrong node and makes clients move to a healthy node. Details in Databases: Fencing.
A limited agent
- The agent only runs a fixed set of commands (install apps, create swap, firewall, databases, load balancers, and so on). It does not run arbitrary shell commands from Saka Panel. The only shell is the emergency terminal that you open yourself with your password.
- MCP and the API use tokens that can be revoked at any time. Delete actions need confirmation, and through MCP they always take two steps. See MCP security principles.
- OpenClaw is installed with command approval on and ClawHub off.
- Sign-up and sign-in are rate-limited per IP to slow down password guessing.
Emergency terminal
Almost everything on a server has a button in the panel. For emergencies, the server page has a Terminal: a root shell on the server through the agent, with no open port and no SSH key.
- Only from the panel by the account owner: not available through API tokens, MCP, or a Saka admin helping you.
- Your account password is asked every time; the terminal disconnects by itself after 15 minutes without typing.
- The Keep the session on the server option (on by default) wraps the shell in tmux, or screen if tmux is missing. Commands keep running even if you refresh the page or your PC shuts down; open the terminal again with your password and your screen continues. Without tmux or screen the terminal still opens without this feature and shows how to install one.
- The session is not recorded. Only who, when, from which IP, and for how long is logged; you get a Telegram message every time the terminal is opened.
- The server owner can turn it off completely: add
"terminal": falseto/etc/sakapanel/agent.json, thensystemctl restart saka-agent.
Account security
- Your email must be verified with a 6-digit code before the panel can be used. The code works for 15 minutes and stops working after 5 wrong tries.
- Forgot password uses a one-time email link valid for 1 hour. After the password changes, all other sessions are signed out and you are notified on Telegram.
- Sign up, sign in, and email requests are rate limited per IP and per account.
- Emails come from
[email protected]with DKIM, SPF, and DMARC.
Team and roles
The account owner can invite others in Settings, Team. Invitees sign in with their own Saka account (the email must match the invitation), so no password is shared.
- Viewers can only look. Developers can deploy, change variables, install sites & apps, back up, and create tunnels, but can't delete. Admins can also delete and manage members.
- Custom roles. If the three roles do not fit, the owner or an admin builds their own in Settings, Team, Custom roles from seven permissions: sites, servers, databases, kubernetes, networking, monitoring, and delete. Viewing is always allowed; changing needs the permission for that area, and deleting also needs delete. Example: a "Site manager" with only the sites permission can deploy and edit sites but cannot touch servers or delete anything. A custom role is chosen when inviting or editing a member, takes effect immediately when edited, and cannot be deleted while in use (maximum 10 per account). A custom role can never manage the team or the owner's personal settings, so there is no way to raise privileges through a role you create. With no permission selected, the role is view-only.
- Permission-scoped API tokens. When creating a token, choose Selected permissions and tick the areas it may change, for example a CI token that may only change sites. The rules match custom roles, and a scoped token cannot manage other tokens, the team, cloud accounts, or the terminal. Through MCP, such a token only sees and uses the tools in the ticked areas: other tools are hidden and refused, tools that delete need the delete permission, and risky actions still ask for Telegram approval. Existing tokens (Can make changes and View only) are unchanged.
- Access can be limited to specific servers: that member only sees those servers and the sites, apps, and tunnels on them. Databases and load balancers are visible only when all their servers are shared.
- Always owner-only: the root terminal, API tokens, AI connectors, Connect to Saka, Telegram, account settings, and S3 recovery codes.
- Alerts on each member's own Telegram. Members who link Telegram in their own account get this account's server alerts (server down, app stopped, site not opening), only for servers they can access; combined alerts about several servers only go to members with access to all servers. Alert buttons follow the role: Viewers do not get Restart, and member actions are recorded under their name and reported to the owner. Members can turn their alerts off on the Team card. The owner sees who has linked Telegram.
- Membership is checked on every request: a removed member loses access immediately. The task history records which member ran each task, and the owner is told when a member joins or deletes something.
Activity log
Account menu, Activity log, records every change in your account: who (you, a team member's email, an API token's name, or an AI agent through MCP with the tool name), when, from which IP, the target (e.g. server or site name), and whether it succeeded. Filter by AI agents, team, API tokens or panel; AI agents can read it with the MCP tool riwayat_aksi. Each row reads as a sentence (e.g. "Created database user analyst"), not an action code; on a phone rows become compact cards and the IP is hidden so they fit. Inputs such as passwords or env variables are not recorded, only the action. Kept for 180 days; visible to the account owner and members with the Admin role.
When the Saka team helps you
If you ask for help, the Saka team can open your panel or run safe actions on your server (read an app's log, restart an app, check a site, create swap). The rules:
- Every time, you are notified on Telegram with the reason.
- Opening your panel is limited to 30 minutes and is view only by default. If something must be changed, every change is logged.
- Your account settings, password, and API tokens cannot be changed by the Saka team.
- Actions on your server show up in the server's history as "via admin". Nothing deletes data.
Stopping using Saka
Saka never stops your apps because a server is disconnected or the agent is removed. To leave cleanly (removing everything Saka installed), use Lepas server (Release server). Details in Release a server.
If Saka Panel goes down
| While Saka Panel cannot be reached | |
|---|---|
| Websites, apps, OpenClaw | Keep running on your server. |
| Databases | Keep running, and still fail over automatically if one server goes down (the primary is elected among your own servers). |
| Load balancers | Keep running. Saka is not in the traffic path or the DNS path. |
| OpenClaw automatic updates | Keep running, handled by the agent itself. |
| Panel, API, MCP | Temporarily unavailable. |
| Telegram notifications | Temporarily not sent. |
| If it stays disconnected for a long time, or the agent is removed | Apps and databases keep running; Saka never stops them. |