Skip to content

SECURITY.md

Latest commit

 

History

History
61 lines (48 loc) · 3.02 KB

SECURITY.md

File metadata and controls

61 lines (48 loc) · 3.02 KB
Raw

Security Policy

Reporting a vulnerability

Open a private security advisory on GitHub, or email vojir.mikulas@gmail.com. Please include what you did, what happened, and what you expected; a proof of concept helps but is not required.

You can expect an acknowledgement within a few days and an assessment within two weeks. Please do not disclose publicly before a fix is released.

Supported versions

Only the latest release is supported. Fixes land on main and ship in the next release rather than being backported.

What RED protects

RED is a desktop database client, so its threat model is mostly about the credentials it holds and the statements it runs on your behalf.

  • Credentials. Database, SSH, proxy and AI passwords live in the operating system keychain, keyed by connection id. They are never written to connections.toml; the shape that file is serialized from has no password field at all. Every file RED writes under the config directory is created 0600 on Unix.
  • Destructive statements. Statements are graded before they run and the riskiest need a typed confirmation. A connection can be marked read-only, and a deployment marker (Local/Dev/Staging/Prod) scales how much RED asks.
  • The AI assistant. Off unless configured, and gated by an access tier (off/schema/read/write) that fails closed: a tool nobody has classified is treated as a write. Every write is refused or shown to you for an explicit approval before it runs; no tool writes without one. The read tier vets each statement rather than trusting the model's own account of it.
  • Transport. TLS is available for every engine, host keys are verified with a trust-on-first-use prompt, and RED never disables certificate verification.

Update trust model

RED checks GitHub for a newer release every six hours (turn it off with auto_update = false under [update] in settings.toml). An update is installed only if it passes every check below, and is refused otherwise:

  • The download host is pinned to github.com, over HTTPS.
  • The version must be strictly newer; a downgrade is refused.
  • macOS: the bundle's signature is verified (codesign --verify --deep --strict), Gatekeeper is asked to assess it (spctl --assess), and its Team ID must match the installed app's. A Team ID that cannot be read is a refusal, not a pass.
  • Linux and Windows: the artefact is checked against the .sha256 sidecar published with the release. These builds are not OS-signed.
  • The new version is staged in a 0700 temporary directory and swapped in atomically, so a failure never leaves a half-written install.

Known limitations

  • The Linux AppImage and the Windows executable are unsigned; the checksum sidecar is what integrity there is.
  • Imported connections from other tools (DBeaver, DBGate, DataGrip, RedisInsight) are decrypted with those tools' own well-known keys. That is what makes the import possible, and it is not a secret RED can keep.