Summary

Audit of the codebase surfaced several high-impact issues. Each is addressed in this PR.

1. Unauthenticated data exfiltration (Critical)

GET /{table}/store, GET /{table}/keys, POST /{table}/keys and the SSE GET /{table}/subscribe/{key} endpoint required no authentication. Any anonymous client could:

curl http://node/auth/store

and walk away with every user's SHA256-hashed password for offline cracking, plus the full contents of every table.

Fix: all four endpoints now require a valid JWT and only return rows whose owner == sub claim.

2. Insecure secret fallbacks (Critical)

• ADMIN_PASSWORD silently defaulted to "admin123".
• CLUSTER_SECRET defaulted to "default_secret" everywhere.
• The admin module had a separate get_jwt_secret() that defaulted to "default_jwt_secret".

Combined with the existing X-Internal-Request: <CLUSTER_SECRET> bypass in get_value, put_value_internal and delete_value, a node that booted without configured secrets would happily read/write/delete any record for any caller using the default header.

Fix: main.rs now refuses to boot unless JWT_SECRET, CLUSTER_SECRET and ADMIN_PASSWORD are all set, ≥16 bytes long, and not in the known-placeholder list ("", "default_secret", "please_change", "admin123", ...). admin_login no longer falls back to a hard-coded password and now uses a constant-time comparison. Admin signing/verification uses the strict engine::get_jwt_secret.

3. Path traversal via {table} / {key} (High)

Table names from URL paths are joined into filesystem paths in persistance.rs (Path::new(state.base_dir).join(&t_name)). A request like

PUT /..%2F..%2Ftmp%2Fevil/key/foo

would write outside ./data.

Fix: new is_safe_name() helper rejects empty names, ., .., names containing /, \, or \0, and names longer than 255 bytes. Applied to every handler (get_value, put_value, patch_value, delete_value, put_value_internal, get_table_store, get_all_keys, get_multiple_keys, subscribe_to_key, admin handlers, authentication::access).

4. Plaintext cluster traffic when DYNA_MODE=https (High)

engine.rs, admin.rs and authentication.rs hard-coded format!("http://{}/...") for all internal replication URLs even when TLS was configured, leaking the cluster secret and every replicated value over the wire.

Fix: network::broadcaster::build_url was already TLS-aware; it is now pub and used by every internal call site.

5. Race condition + ownership forgery in user registration (High)

authentication::access did a non-atomic auth_table.get(&username) then auth_table.insert(...), so two concurrent registrations of the same username could race past each other. The internal replication branch also trusted a caller-controlled User: header as the record owner.

Fix: registration is now atomic via DashMap's entry().and_modify(...).or_insert_with(...). The owner is always the path-derived username; the spoofable User: header is gone. Passwords are now salted with the username before SHA256 to defeat rainbow tables.

⚠️ Breaking change: because the hash function changed (salted vs. unsalted SHA256), existing accounts cannot log in after upgrade and must be re-created. Clear the auth table before deploying, or coordinate a migration.

6. Stored XSS in admin dashboard (Moderate)

admin.html rendered record.owner and vector_clock keys with innerHTML. Any user who PUT a record with an owner containing <img src=x onerror=...> could pop JS in the admin's browser the next time they viewed the record.

Fix: replaced the template-literal innerHTML block with safe DOM construction (textContent + createElement + appendChild).

7. Committed weak credentials (Moderate)

.env was tracked in the repo with JWT_SECRET=123, ADMIN_PASSWORD=admin123, CLUSTER_SECRET=please_change.

Fix: removed .env, added it to .gitignore, and added a .env.example with guidance. Anyone who has ever deployed using those values should rotate them immediately — they are now in the public git history.

Test plan

• cargo build succeeds with no warnings.
• Boot a node with no env file → process exits with a clear fatal: ... refusing to start message.
• Boot a node with JWT_SECRET=123 → process exits (too short).
• Boot a node with CLUSTER_SECRET=default_secret → process exits (forbidden value).
• Boot two properly-configured nodes; verify replication still works and that internal URLs follow DYNA_MODE.
• GET /auth/store without a token → 401.
• GET /default/store with user A's token → only returns rows where owner == A.
• PUT /..%2Fetc/key/foo → 400 Invalid table or key name, no file written outside ./data.
• Register a user; replay the request concurrently → exactly one record, no inconsistent state.
• Log in as admin; PUT a record with owner = "<img src=x onerror=alert(1)>"; reopen it in the dashboard → renders as text, no alert.

https://claude.ai/code/session_01NYckTQH6APLpfGKVwwcQVD

Generated by Claude Code