Client-side validators (reset-password, invite-accept, first-admin setup,
user-create modal) previously checked password.length < 8 while every
server-side Zod schema required .min(12). External API consumers (or a
confused browser UI) could get past the client check but fail at the tRPC
boundary — or worse, quietly under-enforce policy compared to what
admins expect.
Fix: introduce PASSWORD_MIN_LENGTH (12) and PASSWORD_MAX_LENGTH (128) in
@capakraken/shared and import them from every pre-submit client validator
and every server Zod schema. Single source of truth; drift becomes a
compile error rather than a security finding.
Also hardens the AUTH_SECRET runtime check: in addition to the existing
placeholder-blacklist, production startup now rejects secrets shorter
than 32 chars OR with Shannon entropy below 3.5 bits/char. That covers
low-entropy-but-long values like "aaaa..." (38 chars, entropy 0) which
would have passed the previous checks.
Documented the rotation process for AUTH_SECRET + POSTGRES_PASSWORD in
docs/security-architecture.md §3.
Verified:
- pnpm test:unit — 396 files / 1922 tests passed
- pnpm --filter @capakraken/web exec tsc --noEmit — clean
- pnpm --filter @capakraken/api exec tsc --noEmit — clean
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Browser code never calls OpenAI/Azure/Gemini directly; all AI traffic is
server-side tRPC. connect-src is now locked to 'self'. Added object-src 'none',
frame-src 'none', media-src 'self', and worker-src 'self' blob:. style-src
keeps 'unsafe-inline' for React + @react-pdf/renderer (documented residual
risk — script-src is nonce-based so CSS injection cannot escalate to JS).
Added three regression tests covering connect-src no-wildcards, object/frame-src
'none', and worker-src scope.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
checkPromptInjection now NFKD-normalises, strips zero-width / combining
chars, and folds common Cyrillic / Greek homoglyphs before matching. 10
documented bypass examples (fullwidth, ZWJ, ZWSP, soft-hyphen, Cyrillic
е/о, combining marks, LRM, BOM) are covered by unit tests. Security
docs explicitly mark the guard as defense-in-depth — real boundary is
per-tool requirePermission.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Hartmut