Skip to content

Security: bl0rb/Weave

SECURITY.md

Security Policy

Reporting a vulnerability

Use GitHub Security Advisories: open the repository's Security tab and choose "Report a vulnerability". The report stays private between you and the maintainer until an advisory is published.

Do not open a public issue or pull request for a suspected vulnerability, and do not attach a proof of concept to one. Issues and pull requests here are public from the moment they are created.

If private reporting is unavailable to you, write to [email protected] and keep the details out of that first message.

This is a single-maintainer project with no support contract behind it. Reports are read and acknowledged; there is no guaranteed response or fix window. If a report stays unanswered longer than you consider reasonable, a reminder to [email protected] is welcome. Please give a fix a chance to land before writing about the issue publicly.

What a useful report contains

Item Detail
Affected service The directory under services/ — one of api, chat, embeddings, ingest, knowledge, reranker, retrieval, runtime, tools — or contracts/, deploy/, scripts/
Version The tag or commit SHA the finding was reproduced against
Reproduction The requests, inputs, or steps in the order you ran them, including the response that shows the effect
Configuration Which deviations from the service's .env.example or from weave.yaml the finding depends on, and in particular which secrets were set and which were left empty
Impact What an attacker gains, and what access they need to begin

A finding that only reproduces under one particular configuration is still worth reporting. Name the configuration.

Supported versions

The repository carries a single branch, main, with tags up to v0.5.6. There are no release branches and no backports. Fixes land on main; report against main or against the commit you tested.

Scope

In scope: the nine services under services/, the cross-service contracts in contracts/, the Compose stack and database initialization in deploy/, and the configuration renderer scripts/weave_config.py.

Out of scope:

  • The development defaults shipped in the .env.example files. Some name themselves in the value — SECRET_KEY=dev-only-insecure-secret-key-do-not-use-in-production (services/knowledge/.env.example; weave-api ships an empty SECRET_KEY and refuses to start with that default) — and the rest are marked by the comment above them: the SQLite DATABASE_URL fallbacks and the fake embedding and LLM providers. A deployment that keeps any of them is misconfigured rather than vulnerable. Values written as an instruction rather than a default (generate-me-with-openssl-rand-hex-32) are placeholders that cannot start a service at all.
  • Anything that presupposes access to deploy/.env or to the environment variables it is rendered from. That file is generated, is ignored by Git, and holds every service token in the stack; the README's "Internal calls and how each one authenticates" table names three of them as effectively master keys.
  • The two documented properties of the shipped Compose stack: every published port binds 0.0.0.0, and there is no TLS anywhere in deploy/. Both are stated in the README and left to the operator. A way past an authentication check on one of those ports is in scope; the port listening is not.
  • Dependency advisories already covered by CI. .github/workflows/pr-ci.yml runs npm audit --audit-level=high against services/ingest/frontend and pip-audit --strict against services/ingest/backend/requirements.txt. Those two audits are the only ones in the workflow; the other services' dependencies are not audited there, and a finding in them is in scope. A demonstrated exploit path through a dependency remains in scope either way.

Security model

Worth knowing before digging in:

  • Identity lives in Ingest. The gateway receives it through a single-use handoff code and keeps no second user directory (ADR-0006).
  • All containers share one flat bridge network. What separates the services is the bearer token each one demands, checked fail-closed — an unset token yields 503 rather than an open door. The credential on each internal hop is tabulated in the README.
  • Three databases share one PostgreSQL cluster (ADR-0004): weave_ingest, weave_knowledge, and weave_api. Ingest, Knowledge, and Weave-API each connect with the same superuser, so what separates them is a database boundary, not a privilege boundary. The one privilege boundary inside PostgreSQL is weave_retrieval_ro, the SELECT-only role through which Retrieval reads the Knowledge chunk store (ADR-0005).
  • Ingest routes the outbound URLs an administrator can set — OIDC discovery, Confluence import, export webhooks, the vision endpoint — through an SSRF-hardened fetcher (services/ingest/backend/app/services/safe_fetch.py) that rejects private addresses unless the specific host is allowlisted. The endpoints the other services call are requested over plain httpx without that check.

For the detail: docs/betrieb.md is the operations guide (in German) — section 5 covers the values that several services must share, section 6 the misconfigurations that fail silently. docs/adr/ holds the architecture decisions, among them ADR-0002 (authentication and authorization), ADR-0003 (secrets), ADR-0005 (the shared read model), and ADR-0006 (federated login).

There aren't any published security advisories