Skip to content

Security: dbdefault/joons-agent-guard

Security

SECURITY.md

Security Policy

Supported versions

Version Supported
0.1.x Yes
older No

Reporting a vulnerability

Report privately via GitHub Security Advisories: on this repository's Security tab, choose Report a vulnerability. Do not open a public issue or pull request for a security problem.

Please include the affected version, a minimal reproduction, and your assessment of the impact. We respond as soon as possible and credit reporters on request.

Security model

  • No caller authentication. owner, actor, and principal are plain strings the library records as given and never verifies. Feed them from a trusted transport (mTLS identity, verified signature), never from client-supplied headers. Anyone who can reach the database can act as anyone.
  • Database file permissions are the caller's job. All state — lock holders, circuit state, request digests, stored responses — lives in one SQLite file. Anyone who can read the file sees everything; anyone who can write it bypasses every guarantee. Restrict the file to the service account that needs it. The library never sets a file mode: the file is created with your process umask (often world-readable 0644 on POSIX), so create it in a private directory or chmod 600 it before first use; on Windows the parent directory's ACL applies.
  • Size limits. Identifiers are limited to 1024 characters (InvalidIdentifier), circuit failure reasons are truncated to 4096 characters, and each stored replay response is limited to 1 MiB of JSON by default (ReplayGuard(..., max_response_bytes=None) removes the cap). NaN and Infinity are rejected because they are not valid JSON.
  • Corrupted or tampered data fails loudly. A stored response that cannot be decoded raises CorruptStorage; its message names the request but never contains the stored bytes.
  • The CLI reads without writing. Read commands open the database with SQLite's read-only mode, so inspecting a crashed service never rewrites the file (a hot journal is reported as an error instead of being rolled back).
  • Holder disclosure is intentional. LockHeld messages and the CLI show the current holder's owner/job_id so operators can act on them.
  • Do not store secrets in responses. ReplayGuard stores responses as JSON text and replays them later, and reserve_outbound stores the fixed fields verbatim. Store hashes or opaque tokens, never credentials or raw bodies.
  • Nothing expires by age. Stuck state (held locks, pending requests, claimed circuit slots) persists until an explicit operator action; plan the procedures that clear it. See docs/guarantees.md.
  • Locks are advisory. They coordinate only processes that use the same database file through this library, and only on a local filesystem — not network drives.

There aren't any published security advisories