# Wave 16 — Urgent security fixes **Items:** `D19`, `D20`, `D21` **Depends on:** nothing. Independent of waves 12 to 15 (see the decision record). **Decision record:** `docs/waves/decisions-2026-10-02.md` **Background:** `docs/reference/security-review-2026-10-01.md` Four tasks, in build order. All server-side. No page markup or stylesheet changes, so the 390px / 1440px screenshot rule does not apply; the browser probes that read files and SOPs are still listed for a run on a machine with Edge. New check for this wave: `tests/security_check.py`. It is self-contained (its own SQLite database, in-process, no browser), so it runs anywhere. --- ### T16.1 — D21: session decode fails closed; PyJWT 2.15.1 - **Items:** `D21` - **Depends on:** nothing (first task) - **Blocks:** T16.4 - **Surface:** `server/` - **Files:** `server/requirements.txt`, `server/auth.py` (`decode_token`), `tests/security_check.py` (new) **Do:** pin `PyJWT==2.15.1`. In `decode_token`, return `None` for an empty or over-long token (over 4096 characters) and for any exception from `jwt.decode`. Require the `exp` and `sub` claims. **Do not:** change token contents, the cookie, or the D18 refresh. **Done when:** - [x] the crafted-cookie request returns 401, not 500 - [x] a valid session still works, and the D18 refresh still reissues - [x] `pip-audit -r server/requirements.txt` reports no PyJWT advisory - [x] smoke test 27/27 --- ### T16.2 — D19: uploaded files download safely - **Items:** `D19` - **Depends on:** nothing - **Blocks:** T16.4 - **Surface:** `server/` - **Files:** `server/files_http.py` (new), `server/app.py` (`get_wp_file`), `tests/security_check.py` **Do:** serve PDF and raster images inline as today. Serve every other type, SVG included, as an attachment with `Content-Security-Policy: sandbox; default-src 'none'`. Send `X-Content-Type-Options: nosniff` from the API too (NGINX already does; the dev server does not). Build `Content-Disposition` with an ASCII fallback and an RFC 5987 `filename*`. Reference code: `docs/reference/architecture-review/proposals/server/files_http.py` (P-6). **Do not:** refuse SVG uploads (decided: accept, download only). Do not change `Cache-Control`. **Done when:** - [x] an SVG downloads as an attachment with the sandbox header - [x] PNG, JPEG and PDF still open inline - [x] `Plan é.pdf` downloads (was 500) - [x] a name with CR/LF produces a valid header - [x] `tests/files_check.py` passes on a machine with Edge (CR-007 offline drawings) --- ### T16.3 — D20: one SOP per project - **Items:** `D20` - **Depends on:** nothing - **Blocks:** T16.4 - **Surface:** `server/` - **Files:** `server/app.py` (`upsert_sop`, `latest_sop`, `project_summary`, `project_sop_team`, `project_qa_group`), `tests/security_check.py` **Do:** the four rules in D20. One helper picks the project's baseline SOP (canonical row if complete, else latest complete) and every baseline reader uses it. **Do not:** delete or rename existing SOP rows. Do not change any response shape. **Done when:** - [x] a `project_user` cannot create a second SOP on a project (400) or post `complete: true` while a complete SOP exists (403) - [x] a `project_user` can still draft and complete the first SOP of a project - [x] a Project Admin can still change the completed SOP - [x] POST without an id lands on `sop__` - [x] a rogue complete SOP made before the fix no longer drives the baseline or the QA and PM/CM recipients - [x] `tests/sample_check.py` and `tests/critical_reopen_check.py` pass on a machine with Edge --- ### T16.4 — D19, D20, D21: verification - **Items:** `D19`, `D20`, `D21` - **Depends on:** T16.1 to T16.3 - **Surface:** verification - **Files:** this file (verification record) **Done when:** - [x] `tests/security_check.py` passes - [x] `docs/reference/security-review/sec_probe.py`: S3 and S12 change as expected; every other line is unchanged - [x] the architecture review `equivalence_check.py` (current branch vs wave 16) is IDENTICAL: the scenario does not use these paths, so it proves the rest of the work-package flow did not move - [x] smoke test 27/27 - [x] the browser probes named in T16.2 and T16.3 are listed for a run on a machine with Edge (not available in the build sandbox) ## Verification record Run 2026-10-02 in the build sandbox (Linux, Python 3.10, no browser), on branch `feat/wave-16-security` at T16.3. | Check | Result | |---|---| | `tests/security_check.py` | **36/36** (D21 6, D19 14, D20 16) | | Same check against `da1325c` with PyJWT 2.13.0 | 20/36. The 16 failures are the findings (D21 3, D19 6, D20 7). The 20 passes are the "still works" checks plus three that already held (over-long cookie 401, non-member 404, SOP move refused by the existing completed-SOP rule) | | `docs/reference/security-review/sec_probe.py`, before vs after | Exactly 4 of 27 lines change: S3 (new complete SOP 200 -> **400**; baseline `sop_rogue` -> **`sop__proj_a`**; QA group attacker -> **real group**) and S12 (crafted cookie 500 -> **401**). All other lines identical | | `equivalence_check.py` (architecture review), `da1325c` vs wave 16 | **IDENTICAL**: 78 HTTP results, 89 audit rows, 24 notifications. The scenario's SOP id was changed from `sop_eq` to the canonical `sop__proj_eq` on both sides, because D20 refuses a non-canonical id on purpose | | `server/smoketest.py` | **27/27** after each of T16.1, T16.2, T16.3 | | `server/seed_demo.py` | Seeds the demo project; its SOP lands on the canonical id | | `pip-audit -r server/requirements.txt` | No known vulnerabilities | **Browser probes, run 2026-10-02 on Matt's machine (Windows, Edge 154, Python 3.12.10, `.venv`), one at a time:** | Probe | Task | Result | |---|---|---| | `tests/files_check.py` | T16.2 | **36/36**, including the 1440px creator, 390px Field View, and the offline drawing cache with the server gone | | `tests/sample_check.py` | T16.3 | **10/10** | | `tests/critical_reopen_check.py` | T16.3 | **11/11**, PM + CM + assignee mail, sink-verified | T16.2 and T16.3 are done. **Production data check for D20, 2026-10-02.** On the production database (`wp_db`, psql): projects with more than one complete SOP: **0 rows**. D20 has no legacy duplicates to handle on deploy. **Harness fix needed to get there (`tests/cdp.py`, committed with this record).** Every probe failed first with "browser would not start ... exited rc=0 without binding the port". Cause: the PowerShell session was elevated, and Edge and Chrome de-elevate by relaunching themselves through Explorer as a normal user; the relaunch drops `--remote-debugging-port`. Edge's log said so (`Edge is running elevated: 1`). No browser policy was involved (all hives checked). Two changes, both inert on a normal shell and on Linux/macOS: `--do-not-de-elevate` on Windows, and a 15-second grace for the port when the launched process exits rc=0. Probe results are unchanged by it; it only lets the browser start.