Do not open a public issue for security problems.
Report privately via GitHub Security Advisories: the Security tab → Report a vulnerability
(or .../security/advisories/new). You will get a private channel for discussion and a coordinated fix.
Include, if you can: the affected version, reproduction steps, the impact, and a proof of concept.
No GitHub account? Email stefan.maldaianu@gmail.com with WebTerm security in the subject.
Advisories are preferred — they give us a private thread and a coordinated fix — but not having an
account should never be the reason a problem goes unreported.
What to expect. The project is maintained by one person, in his spare time, so this is a promise about honesty rather than speed: an acknowledgement within 7 days, and an assessment (in scope or not, and roughly when) within 30 days. If you have not heard back in 30 days, assume the message was lost and ping again — that is a failure on my side, not impatience on yours. Fixes ship in the next release; you decide whether to be credited in the advisory.
Safe harbour. Test only against your own installation. Do not access other people's data, do not degrade a service someone else depends on, and give me a reasonable window before publishing. Within those limits I will not pursue anything against a good-faith report, and I would rather hear about a finding that turns out to be nothing than not hear about one that was real.
Bugs that break the documented trust boundaries (see docs/THREAT-MODEL.md):
- authentication / session / lockout bypass;
- Cross-Site WebSocket Hijacking, CSRF, XSS, injection (SQL/ANSI/OSC/header);
- path traversal, SSRF, secret leakage;
- accepting unsigned / downgraded agent updates, cert pinning bypass;
- forward isolation (token bound to slug), cross-host escalation;
- roles: reaching a host outside your scope, getting a shell without a shell permission, raising your own role, or removing the last Owner;
- sharing: a live share or replay link that outlives, or reaches beyond, the access of the account that created it.
WebTerm has several accounts with roles: the Owners are fully trusted, and every other account is trusted exactly as far as its role reaches (see docs/ROLES.md). The following are deliberate decisions, explained in docs/THREAT-MODEL.md, not bugs:
- The agent runs with its user's privileges on every host and never escalates; fs/command
operations are NOT sandboxed. Anyone who gets past login has that user's access across the whole
fleet — like holding an SSH key for it. The install command offered by default creates a dedicated
unprivileged
webtermuser; installing as the current user while root makes it root's. - Roles separate hosts, not people sharing a shell. Since 3.6.0 accounts have roles (Owner / Admin / Operator / Viewer over all hosts, a folder, a tag or one host — see docs/ROLES.md). They enforce which hosts an account can reach and whether it gets a shell there. Inside a host where an account holds a shell-equivalent permission it has the agent user's full power, and a shell on the host that runs WebTerm (or holds its backups) is equivalent to Owner. A role-check bypass — reaching a host outside your scope, getting a shell without a shell permission, or escalating your own role — is a vulnerability; please report it.
- The gateway is a single point of total compromise (it commands the agents). A compromised gateway = a compromised fleet.
- The gateway host holds the vault key, and so do its backups' secrets. Scheduled backups
(
scripts/backup.sh, thewebterm-backuptimer) are encrypted — AES-256-CBC with PBKDF2 viaopenssl enc, passphrase in/etc/default/webterm-backup, generated byinstall.sh— and a non-interactive run without a passphrase deletes the archive and exits 1 rather than write the vault key in cleartext (WEBTERM_BACKUP_ALLOW_PLAINTEXT=1is the explicit opt-out). The download from Settings → Backup uses a password of your choosing (scrypt → AES-GCM,WTBK1). The limitation is where the passphrase lives: on the same host as the data. Root on the gateway host can read every credential regardless of backup encryption — that is the premise above, not a weakness of the archive format. - The
insecuremode (TLS without validation) is an opt-in, documented footgun.
If you find a way to break one of these without the premises above (e.g. escalation from unauthenticated, from one account into another's sessions, files or links, or beyond what an account's role grants), that is in scope. Please report it.
Policy last reviewed: 2026-11-18 (3.6.36).
Security fixes go into the latest release only; older versions get no backports. Upgrade to it.
| Channel | What it is | Supported |
|---|---|---|
ghcr.io/sm26449/webterm:vX.Y.Z (latest tag) |
A release: the full test gate ran on it, and the image is signed (cosign keyless, with provenance and SBOM). The default of install.sh and .env.prod.example. |
Yes, the latest one |
ghcr.io/sm26449/webterm:latest |
The tip of main: tested by CI, but not released. upgrade.sh never falls back to it unless you pass --allow-latest. |
Best effort, for testing |
| Building from source | Your own build of a tag or of main. |
As the tag / main it came from |
Older vX.Y.Z tags |
Previous releases. | No |
Report against the version you run. If you run latest or a build, include the commit.
- Run with a domain + HTTPS + passkeys, not just IP/password.
- Install agents as a dedicated least-privilege user, not root, wherever you can.
- Your agent signing key is generated automatically on the first boot of an install with no
hosts (unencrypted, so auto-updates survive a restart); Settings → Infrastructure & tokens shows it. To use your
own or a passphrase-protected key instead, set it up before first boot or before enrolling
hosts: generate/import return 409 once a key exists, so you place
data/agent-signing.key(PEM, mode 600) anddata/agent-signing.pub(its public key, hex) yourself and restart the gateway. Swapping it after hosts are enrolled means reinstalling every agent (or the rollover in docs/design/SIGNED-UPDATES.md). - Do not expose
insecuremode in production.