From 1aa4ca68c823c73df7acfcf80862df95e836b833 Mon Sep 17 00:00:00 2001 From: terra tauri Date: Sun, 27 Sep 2026 01:21:07 -0700 Subject: [PATCH] docs(security): record the multi-replica replay-guard residual MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The optional replay guard is per-process, so it protects a single backend instance. Behind more than one replica the in-memory guard gives false comfort; a horizontally scaled backend needs a shared store behind the ReplayGuard interface. State that plainly, and why no shared implementation ships here (it would force a datastore choice onto the generic library). Documented gap, not a hole — the deployed StatefulSet runs one replica. Co-Authored-By: Claude Opus 4.8 Claude-Session: https://claude.ai/code/session_01SDfMmeRXGsanqzDDSjL1uU --- SECURITY.md | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/SECURITY.md b/SECURITY.md index ddd5505..fcacc6a 100644 --- a/SECURITY.md +++ b/SECURITY.md @@ -56,6 +56,13 @@ single-use; both are weaker than confining the hop. Restrict who reaches the backend (a network policy), and keep the lifetime short. This is the same trust that every identity-aware proxy places in the network behind it. +The shipped replay guard is per-process: it protects one backend instance. A +horizontally scaled backend needs a shared store behind the `ReplayGuard` +interface (the interface exists so one can be plugged in); running more than +one replica with the in-memory guard gives false comfort, not protection. No +shared implementation ships here on purpose — it would pull in a datastore the +generic library has no business choosing. + ## What is not in scope * **A stolen device.** A person with an unlocked device gets the identity of