Vulnerability Disclosure Policy
This page is the full coordinated vulnerability disclosure (CVD) programme for
LumiBase. The short version lives in SECURITY.md at the
repository root; where the two differ, SECURITY.md is fixed to match this page.
Related pages:
- Security advisories — where advisories are published and how to subscribe.
- Advisory and CVE post template — the text maintainers publish.
- Advisory runbook — the maintainer procedure, with the dry run of the first advisory.
1. How to report
Report privately through GitHub private vulnerability reporting: the repository's Security tab → Report a vulnerability, or directly https://github.com/khuepm/lumibase/security/advisories/new. The report becomes a draft GitHub Security Advisory (GHSA) that only you and the maintainers can see; all discussion, the fix and the CVE request happen inside it.
Do not report through public issues, discussions, pull requests or social media.
The issue chooser in .github/ISSUE_TEMPLATE/config.yml links to the private form.
If you cannot use GitHub, open a public issue titled "Security contact request"
with no details and a maintainer will open a private channel.
A complete report contains:
| Field | What we need |
|---|---|
| Affected version | Release tag (v2.0.0) or commit SHA. Must be a supported version or main. |
| Component | CMS on Cloudflare Workers, CMS Docker image, Studio, desktop/mobile shell, or an npm package (name and version). |
| Configuration | Anything you changed from the defaults (environment variables, feature flags, deployment profile). |
| Reproduction | Numbered steps from a clean install (docker compose -f docker/docker-compose.yml up -d or a fresh Worker deploy). |
| Proof of concept | A working request sequence, script or failing test, with the output you observed and the output you expected. |
| Impact | Which trust boundary is crossed (tenant, organization, role, approval gate, secret) and what an attacker gains. |
| Credit | The name or handle to credit, or "anonymous". |
2. Supported versions
Fixes follow the support window in the versioning policy: the current major, plus the preceding major for 6 months after a new major ships; within a major only the latest minor.
| Version | Released | Supported |
|---|---|---|
2.0.x (current major) | 2026-11-10 | Yes — fixes ship as 2.0.x patch releases |
1.0.x (preceding major) | 2026-11-09 | Security fixes only, until 2027-05-10 |
1.0.0-rc.* pre-releases | 2026-09 | No |
0.x | before 2026-09 | No |
A report against an unsupported version is accepted only if the PoC also works on a supported version. This table is updated by the release that changes it (see runbook §8).
3. Scope
In scope is anything that lets an actor cross a trust boundary in a default or
documented configuration: tenant or organization isolation, authentication and
authorization (sessions, tokens, RBAC, the AI approval flow), injection (SQL,
path traversal, SSRF, template or expression injection), secret exposure,
prompt-injection paths that reach schema:write or destructive skills without an
ai_approvals gate, and the supply chain of what we release (the Docker image
built by .github/workflows/release.yml and the published npm packages).
Out of scope: issues that need a pre-compromised host or an admin acting inside
their own tenant; configurations you made insecure yourself; denial of service
by unrealistic volume against an unprotected dev node; dependency CVEs with no
reachable path in LumiBase; and unverified scanner or AI output (see
§8). The full list is in
SECURITY.md.
4. Severity
Severity is the CVSS base score of the vulnerability in its default or documented configuration. We use the CVSS version that GitHub's advisory form offers and publish the vector string with the advisory.
| Severity | CVSS | LumiBase examples |
|---|---|---|
| Critical | 9.0–10.0 | Unauthenticated remote code execution; authentication bypass to any site; reading JWT_SECRET or ENCRYPTION_KEY remotely. |
| High | 7.0–8.9 | Cross-tenant or cross-organization read/write by an authenticated user; privilege escalation to site admin; an agent reaching schema:write without an approval. |
| Medium | 4.0–6.9 | Stored XSS in Studio that needs an editor role; SSRF limited to non-internal hosts; leaking another tenant's metadata (names, counts). |
| Low | 0.1–3.9 | Information disclosure without tenant data; issues that need an unlikely configuration or user interaction. |
Two rules adjust the score. A finding that crosses tenant or organization isolation is at least High, because multi-tenancy is the product's core promise. A finding that only works in a non-default configuration is scored for that configuration, and the advisory says so.
5. Response targets
Clocks run in calendar days unless marked as business days. The triage clock starts when the report is complete (§1).
| Step | Target |
|---|---|
| Acknowledgement | 3 business days after the report |
| Triage: in scope or not, severity, affected versions | 7 business days after a complete report |
| Fix released — Critical | 14 days after triage |
| Fix released — High | 30 days after triage |
| Fix released — Medium | 60 days after triage |
| Fix released — Low | 90 days after triage, or the next minor release |
| Public advisory | When the fix ships; at most 90 days after the report unless extended (§7) |
If a target will be missed we say so in the advisory thread before it passes, with the reason and a new date.
6. Advisories and CVE IDs
Every confirmed vulnerability in a released version gets a GitHub Security Advisory in this repository. Maintainers request a CVE ID from GitHub, which is a CVE Numbering Authority, for every advisory of Medium severity or higher; a Low advisory gets one on request. The CVE is assigned to the draft and becomes public when the advisory is published.
Published advisories appear at
https://github.com/khuepm/lumibase/security/advisories, in the GitHub Advisory
Database for the npm packages (which drives Dependabot alerts and npm audit),
and under ### Security in the release's CHANGELOG.md entry. Details are on the
security advisories page. A vulnerability found only on main
and fixed before any release contains it gets no advisory, only a ### Security
changelog line.
7. Embargo and credit
- Embargo. Keep details private until the advisory is published or 90 days have passed since your report, whichever is first. A complex fix can need more time: we ask for an extension in the advisory thread, give the reason and a date, and do not extend without your agreement. If a vulnerability is being exploited in the wild, or becomes public some other way, we may publish early with mitigations and ship the fix as soon as it is ready.
- Proof-of-concept details are published with the advisory or up to 14 days later, so operators can upgrade first.
- Credit. Reporters are credited in the advisory's credits field (as
"reporter", plus "remediation developer" if you contributed the fix) and in the
CHANGELOG.mdline, by the name or handle you choose. Anonymous is fine. We do not run a paid bug bounty. - Collaboration. We can invite you to the advisory's temporary private fork to review or write the fix.
8. AI-generated and automated reports
Using AI tools or scanners to find a vulnerability is fine. Sending their output unverified is not a report. The person who submits is responsible for having reproduced every claim.
Every report, however it was produced, must have:
- A supported version (§2) named by tag or commit, and the PoC run against it.
- Reproduction steps a maintainer can follow from a clean install.
- A working PoC — the exact requests, script or failing test — with the observed output. A description of what a PoC "could" do is not a PoC.
- A concrete impact on a boundary listed in §3, shown by the PoC rather than argued.
- References to code that exists: file paths, functions and routes as they are in that version.
Not accepted as a report:
- Raw output of a scanner, SAST tool or LLM, or a list of CWE matches.
- "Could lead to", "may allow" or "potentially" claims with no demonstrated path.
- Findings in test files, fixtures, examples, dev scripts or other code not shipped in a release.
- Dependency CVEs without a reachable path in LumiBase (Dependabot covers those).
- The same finding resubmitted with different wording after it was closed.
How such reports are closed. The triager timeboxes reproduction to about 30 minutes.
| Situation | Action |
|---|---|
| PoC missing or does not reproduce, but the claim is plausible | Reply with the "needs PoC" message from the template. The reporter has 14 days to supply one; without it the report is closed. |
| Cited files, functions, routes or settings do not exist in LumiBase | Closed as invalid at once, with one line naming what does not exist. |
| Scanner or LLM dump, or impact not on an in-scope boundary | Closed with the "not a vulnerability report" message. |
| Plausible, but only against an unsupported version | Closed, asking for a PoC on a supported version. |
Closing means closing the draft advisory on GitHub, so nothing is published. Reports from a reporter whose previous reports were repeatedly closed as invalid are triaged after all other open reports. A closed report is reopened if a working PoC arrives later.
Telling us that AI tools helped is welcome and never counts against a report. Well-reproduced AI-assisted reports are handled and credited like any other.
9. Safe harbour
We will not take legal action against, or ask a host to act against, research that follows this policy in good faith:
- Test only against instances you run yourself. Do not test
lumibase.dev,docs.lumibase.devor anyone else's deployment. - Do not access, change or delete data that is not yours. Stop when you can show impact, and report.
- Do not degrade service for others (no volumetric denial of service, no spam).
- Keep the details private under the embargo in §7.
If you are unsure whether something is allowed, ask in the advisory thread first.