# REPORT — Security review, round 2 (2026-11-19) A second, independent reader of our own code. It looked and wrote down; it changed no product code, no box, no hub, no key. ## In plain language - **What I found: 29 findings.** 1 critical, 10 high, 14 medium, 3 groups of small ones (about 32 small items), and one entry that repeats two open October rows. - **New, known, came back:** 26 are new. 3 were known (one of them was ranked too low). **None came back after a fix.** All of October's fixes that I could test still hold, and their tests pass. - **P1 (usable by a stranger today): none confirmed. One is open as a question.** The scrambled form of the hub's first password has been public since February. If that password is still the live one, it is a P1. I could not read which password is live. **You answer this (D91).** - **The critical one:** a controller that someone has taken over can very likely become root on the host machine. It needs the controller first, so a stranger cannot use it alone. - **The pattern in the highs:** the check that stops a bad app definition (built after October) can be passed in six ways — by an imported app bundle, and by a catalog change alone. - **October findings this pass missed:** 3 small items of one October row; and 1 October finding I did not re-read. Details: `documentation/audits/security-review-2-2026-11-19/COMPARISON.md` §4. - **You decide two things** (`STATUS.md`, top of the decide table): **D91** — change the hub password now (my pick: yes); **D90** — close or accept that a taken-over hub can take over dashboards (my pick: close). - **Not reached:** most of the controller's backup code; the large root helper for OS updates beyond its entry checks; the hub's live settings (ingress, who may push to Gitea); the public sites beyond their headers. The full list is in `COVERAGE.md`, last column. - **Rows:** 254 before, 272 after, 18 opened, 0 closed. Noted: R-984 (re-ranked P4 → P2), R-979, R-980, R-974, R-955, R-1054. ## What happens next Project Claude checks the findings and writes the fix tasks. Nothing was fixed here, on purpose. ## Baselines | repo | reviewed commit | October's commit | |---|---|---| | `felhom.eu` (hub 0.182.0 deployed) | `b7d8f75b` | `0d463a85` (October's own page says `c9b94329`; the fix was `b67b8cea`) | | `felhom-controller` (0.353.0 on demo-hp) | `10df99b8` | `dcd89364` | | `felhom-agent` (0.175.1 on demo-hp) | `3411592b` | `86adbcaf` | | `app-catalog-felhom.eu` | `b6058150` | `1a642b43` | | `homelab-manifests` (clone at `~/git/homelab-manifests`) | `1908bc68` | — | **Correction to this report's first version:** the agent commit was written `34115920`; the right short hash is `3411592b`. The review read a `git archive` export of each commit, not the working tree (another session worked in the same clone all morning). ## Repository visibility (anonymous GET from DooPlex, 2026-11-19 08:40 CET) `felhom.eu`, `felhom-controller`, `felhom-agent`, `app-catalog-felhom.eu`: page 200, clone 200. `homelab-manifests`: 404. Registry: 401 on `/v2/`, the catalog and a hub image manifest. So the attacker has read all product source (accepted, R-925). Limit: the request left DooPlex; no request was made from outside the LAN. ## Method, and what was measured - Seven read-only helpers, one area each, with the brief's fences in full; each returned leads. Six fresh second readers re-traced every Critical and High without seeing the reasoning: **all hold, none disputed.** - Live, only on `demo-hp` and scratch guest 9202 (endpoint and shell reads; no browser): listeners, a full TCP connect scan of the host's LAN address (`nmap` is not on DooPlex; a plain connect loop), sshd, sudoers, file permissions, guest settings; three facts behind S2-01 (read-only, one of them in a private mount namespace that left nothing); three compose and Docker behaviours behind S2-06, S2-08, S2-09 with throwaway containers. - **Teardown, three layers:** machine — the throwaway containers and temp folders on 9202 were removed in the same command (0 left, checked); host — nothing left (0 mounts, 0 temp folders, checked); hub — provisioned nothing. - **One act was refused to CC and not done:** a read of the running hub (`kubectl exec`) to see which operator password is live. Tried: that one command. So S2-10 is „static, live state not measured". - **One secret-derived value appeared on a helper's screen by accident:** the operator password's bcrypt hash, when a search printed line 74 of the public hub manifest. It was not copied anywhere. - Evidence (never committed): `/mnt/5_hdd/felhom.eu/evidence/security-review-2-2026-11-19-PRIVATE/` (0700): `leads/`, `second-read/`, `live/`, `INDEX.md`, and the source exports. ## Commits (felhom.eu `main`, documents only) `5f8412a8` Phase 0 · `3e7c0502` (rebased) five rows + R-984 · this commit: findings, coverage, comparison, 13 rows, STATUS, REPORT. ## Instruction files and stale facts No `CLAUDE.md` or rule file was edited. Stale facts seen, **not** corrected here (they sit in files the release session is editing today or in architecture documents — for the fix tasks): - `13-threat-model.md` §5 „guest ↔ Proxmox host: holds" and „agent ↔ root on the host: holds by the sudoers" — S2-01, S2-23. - `13` §4 „+8 in setup mode" — setup mode is gone. `13` §4 „felhom-sshd (tunnel only)" — true, measured. - `OPEN-ITEMS.md` section headings: five carried a wrong row count before this session (for example „Backup & restore — 49 rows", 52 by count). **Corrected** by a recount of every heading (file `OPEN-ITEMS.md`, the `## …` lines; no gate reads them). ## System poster No fact of `felhom-system-poster.facts.md` changed (no machine, role, path, tier, time, retention or key). Nothing done on the three. ## Observations 1. The allow-list is asked on every catalog path but not on the import path, and it judges one file. FILED: R-1407, R-1409, R-1420. 2. A controller-to-agent route trusts a caller-named path as root. FILED: R-1406; backup target: R-984 (note, re-ranked). 3. Page scripts insert server text unescaped in about 20 places; four are proven reachable. FILED: R-1408. 4. A bind is judged at write time only. FILED: R-1421. 5. Hub sign-in: public start hash, no try limit, no sign-out. FILED: R-1410 (D91). 6. The hub's unsigned channels to the controller. FILED: R-1422 (D90). 7. An unsigned job crashes the agent. FILED: R-1423. 8. The hub and the box relay as open mailers for a box. FILED: R-1424. 9. Hub hand-over doors (re-enroll, appliance register, escrow replace). FILED: R-1425. 10. No per-box limits on the hub; ingress settings unknown; ring 0 by one box. FILED: R-1426. 11. Samba path, storage path, claim page lock. FILED: R-1427. 12. Root helpers and the format confirmation. FILED: R-1428. 13. One CI runner for four repositories. FILED: R-1429. 14. Low items, three groups. FILED: R-1430, R-1431, R-1432. 15. Controller image by tag; installer by a movable tag. FILED: R-979 (a dated note there, and one on R-980). 16. `settings.json` and the debug ring are world-readable inside the guest. NOT-A-FINDING: only root exists in the guest; the file holds bcrypt hashes and one encrypted value (measured). October listed it (R-984). 17. traefik holds the Docker socket read-only-mounted. NOT-A-FINDING: needed by its provider; known design; it is why S2-08 matters. 18. `cf_tunnel_token` appears 215 times in a week of controller log. NOT-A-FINDING: it is the key's name in a „no token" line; no value (measured). 19. Host root filesystem's id is mountable through the assign route too. NOT-A-FINDING: it is the same missing check as R-1406, and it is in that row's private notes. 20. A host with a scratch guest beside the customer's gives each guest's token every host-wide route. NOT-A-FINDING: test boxes only; the design is one customer per host. 21. Vulnerable dependencies seen on the way. NOT-A-FINDING: none was seen (FileBrowser 1.5.8 on demo-hp is above the fix version of R-1305). 22. The controller's `docker-compose.yml` in the repository still shows `privileged` and port 8080; demo-hp's running controller is not privileged and publishes no port (measured). FILED: R-1431 ## Minutes per phase Phase 0: about 10. Phases 1–3 (areas read in parallel by helpers, second reads, live measurements): about 30. Phase 4 (comparison, rows, pages): about 15.