Sicherheit: Tomcat 11.0.25, Jackson 3.1.7, postgresql 42.7.14, Workflow-Permissions

.github/workflows/ai-testgen.yml

@@ -16,6 +16,12 @@ on:
   pull_request:
   workflow_dispatch:
 
+# Least privilege (CodeQL actions/missing-workflow-permissions, 2026-10-21): the jobs
+# here only read the repository; artefact upload and the Actions cache use the
+# runtime token, not GITHUB_TOKEN, so read access to the contents is all it needs.
+permissions:
+  contents: read
+
 jobs:
   harness:
     runs-on: ubuntu-latest

.github/workflows/ci.yml

@@ -11,6 +11,12 @@ concurrency:
   group: ci-${{ github.ref }}
   cancel-in-progress: true
 
+# Least privilege (CodeQL actions/missing-workflow-permissions, 2026-10-21): the jobs
+# here only read the repository; artefact upload and the Actions cache use the
+# runtime token, not GITHUB_TOKEN, so read access to the contents is all it needs.
+permissions:
+  contents: read
+
 jobs:
   test:
     runs-on: ubuntu-latest

.github/workflows/e2e.yml

@@ -13,6 +13,12 @@ on:
     - cron: '0 2 * * *' # nightly, 02:00 UTC
   workflow_dispatch:
 
+# Least privilege (CodeQL actions/missing-workflow-permissions, 2026-10-21): the jobs
+# here only read the repository; artefact upload and the Actions cache use the
+# runtime token, not GITHUB_TOKEN, so read access to the contents is all it needs.
+permissions:
+  contents: read
+
 jobs:
   e2e:
     runs-on: ubuntu-latest

.github/workflows/legacy-ci.yml

@@ -11,6 +11,12 @@ on:
   pull_request:
   workflow_dispatch:
 
+# Least privilege (CodeQL actions/missing-workflow-permissions, 2026-10-21): the jobs
+# here only read the repository; artefact upload and the Actions cache use the
+# runtime token, not GITHUB_TOKEN, so read access to the contents is all it needs.
+permissions:
+  contents: read
+
 jobs:
   legacy-build:
     runs-on: ubuntu-latest

.github/workflows/modern-ci.yml

@@ -11,6 +11,12 @@ on:
   pull_request:
   workflow_dispatch:
 
+# Least privilege (CodeQL actions/missing-workflow-permissions, 2026-10-21): the jobs
+# here only read the repository; artefact upload and the Actions cache use the
+# runtime token, not GITHUB_TOKEN, so read access to the contents is all it needs.
+permissions:
+  contents: read
+
 jobs:
   modern-build:
     runs-on: ubuntu-latest

.github/workflows/playbook.yml

@@ -22,6 +22,12 @@ on:
       - '.github/workflows/playbook.yml'
   workflow_dispatch:
 
+# Least privilege (CodeQL actions/missing-workflow-permissions, 2026-10-21): the jobs
+# here only read the repository; artefact upload and the Actions cache use the
+# runtime token, not GITHUB_TOKEN, so read access to the contents is all it needs.
+permissions:
+  contents: read
+
 jobs:
   playbook-pdf:
     runs-on: ubuntu-latest

ai-testgen/testbed/modern/pom.xml

@@ -73,7 +73,7 @@
 		<dependency>
 			<groupId>org.postgresql</groupId>
 			<artifactId>postgresql</artifactId>
-			<version>42.7.12</version>
+			<version>42.7.14</version>
 		</dependency>
 		<!-- Added 2026-08-05 because stage 6 added them to modern/pom.xml, and the code
 		     under test is whatever that module is today. PomDriftGuardTest caught the

docs/DEVIATIONS.md

@@ -15,7 +15,7 @@ Status values: **deferred(stage)** = lands in a named stage · **deferred(post-v
 | Standard (DoD) | Status | Disposition |
 |---|---|---|
 | Coverage-Gate ≥ 80 % auf Neucode | **met** (2026-08-02, G6) — line only, branch still open | Armed as a **ratchet** throughout, always just below a *measured* value and then pulled up. 2026-07-31: the module's first suite reached 37.3 % line / 14.2 % branch → gate 0.35. 2026-08-02: the repaired G6 corpus-B tests were adopted into `modern/src/test` under [ADR-0011](adr/0011-adopting-generated-tests.md) — 88 test methods, 99 green in the module — and coverage measured **81.3 % line / 58.3 % branch** → gate raised to **0.80**. That is the first time §3's 80 % is *reached* rather than declared. **Two things stay open and are not hidden by the number:** the gate limits **lines only** (branch coverage lags 23 points behind and gets a limit in the same pass that improves the God-class suite), and the adopted `WerkstattService` suite has a **44.1 % mutation score** — high line coverage over the riskiest class with half the injected faults surviving. Coverage went up; the God class is still the weak spot, by measurement. |
-| Security-Scan in CI | **met** (2026-08-05, G7) — image scanning is **reporting-only**, stated not hidden | Since 2026-07-30: Dependabot (modern/e2e/characterization + actions; since 2026-07-31 also **npm for `modern/frontend`** — stage 5 had opened the ecosystem without scanning it, caught by review session 10; `legacy/` **excluded on purpose** — its EOL dependencies are the exhibit). 2026-08-05: **Trivy** scans the built modern image in `modern-ci` for CRITICAL/HIGH in OS packages and Java dependencies — the layer Dependabot structurally cannot see, because a CVE in the JRE base image appears in no manifest and ships anyway. Two limits belong next to that: it runs with `exit-code: 0` (**reporting, not blocking** — a fresh CRITICAL in a base image arrives without any commit of ours, and a gate that reddens master overnight for something nobody caused is a gate people learn to bypass), and a separate OWASP dependency-check was **not** added, because it would duplicate what Dependabot and Trivy already cover and now needs an NVD API key to run at usable speed. **It paid for itself on the first CI run:** one HIGH finding, **CVE-2026-54291** in `org.postgresql:postgresql` 42.7.11 — a man-in-the-middle protection bypass via a SCRAM-SHA-256-PLUS downgrade, i.e. precisely the authentication mechanism PostgreSQL 18 uses, which stage 6 had just switched the modern stand to. Fixed the same day by pinning 42.7.12 (the minimum fixed version, not the newest — this is a security fix, not a version bump). Fixing it exposed a second, quieter hole: the G6 testbed and `modern/pom.xml` both left the driver version to Boot's dependency management, and a property override travels through Boot's `<parent>` but **not** through the testbed's imported BOM — so `PomDriftGuardTest` stayed green while the two classpaths silently disagreed. Both now declare the version literally, which is what makes the guard able to see it. After the pin the scan reports **0 findings** across the base layer, the jar and the Go binary. |
+| Security-Scan in CI | **met** (2026-08-05, G7) — image scanning is **reporting-only**, stated not hidden | Since 2026-07-30: Dependabot (modern/e2e/characterization + actions; since 2026-07-31 also **npm for `modern/frontend`** — stage 5 had opened the ecosystem without scanning it, caught by review session 10; `legacy/` **excluded on purpose** — its EOL dependencies are the exhibit). 2026-08-05: **Trivy** scans the built modern image in `modern-ci` for CRITICAL/HIGH in OS packages and Java dependencies — the layer Dependabot structurally cannot see, because a CVE in the JRE base image appears in no manifest and ships anyway. Two limits belong next to that: it runs with `exit-code: 0` (**reporting, not blocking** — a fresh CRITICAL in a base image arrives without any commit of ours, and a gate that reddens master overnight for something nobody caused is a gate people learn to bypass), and a separate OWASP dependency-check was **not** added, because it would duplicate what Dependabot and Trivy already cover and now needs an NVD API key to run at usable speed. **It paid for itself on the first CI run:** one HIGH finding, **CVE-2026-54291** in `org.postgresql:postgresql` 42.7.11 — a man-in-the-middle protection bypass via a SCRAM-SHA-256-PLUS downgrade, i.e. precisely the authentication mechanism PostgreSQL 18 uses, which stage 6 had just switched the modern stand to. Fixed the same day by pinning 42.7.12 (the minimum fixed version, not the newest — this is a security fix, not a version bump). Fixing it exposed a second, quieter hole: the G6 testbed and `modern/pom.xml` both left the driver version to Boot's dependency management, and a property override travels through Boot's `<parent>` but **not** through the testbed's imported BOM — so `PomDriftGuardTest` stayed green while the two classpaths silently disagreed. Both now declare the version literally, which is what makes the guard able to see it. After the pin the scan reports **0 findings** across the base layer, the jar and the Go binary. *2026-10-21:* the same scan had climbed back to **12 fixable CRITICAL/HIGH** — 8 in the jar (Tomcat 11.0.24: 3 CRITICAL; Jackson 3.1.5: 5 HIGH) and 4 in a Go binary — none caused by a commit of ours, which is exactly the overnight drift the reporting-only decision anticipated. Fixed by pinning the minimum fixed versions over Boot 4.1.1's (Tomcat 11.0.25, Jackson 3.1.7) plus postgresql 42.7.14. **Measured after the fix: 4 HIGH, 0 CRITICAL, jar 0.** All four are Go standard-library CVEs in `/usr/bin/pebble` — Canonical's service manager, shipped by the `eclipse-temurin:25-jre` base image (Ubuntu 26.04), built with Go 1.27.1 (fixed in 1.27.3). Our entrypoint never starts it. Not removed in our Dockerfile: deleting a base-image file in a later layer hides it from the scanner without removing it from the image, which would turn an honest count into a cosmetic one; it closes when the base image is rebuilt. The current count is also on `/nachweis`. |
 | AuthN/AuthZ (OAuth2/OIDC, Keycloak) — §4 | **G7's hard requirement met** (2026-08-05) at the edge · full OIDC still **owner re-scoping needed** | The 2016 app has no auth at all; adding it changes every pinned contract, so it cannot happen as a side effect (ADR-0004 gate). Recorded honestly instead of silently (review session 10): the modern stand — including the destructive `POST /admin/bereinigen` the JSP wart B16 bequeathed — is unauthenticated by inherited design. The G7 requirement was *"protect `/admin` at minimum (reverse-proxy auth counts)"*, and that is exactly what shipped: Basic auth in a Traefik overlay covering `/admin`, `/api/admin` and `/actuator` ([ADR-0014](adr/0014-authentifizierung-am-edge.md)), measured 401 without credentials and 200 with them, wrong password 401. **What that is not:** the application itself is still unauthenticated, so anyone who can reach its port bypasses the lock — which is why a real host must not publish it (`deployment.md` §12). The measured price of doing it in the application instead: a matcher scoped to `/admin` alone turns 4 safety-net tests red, each one an ADR-0004 sanctioned-divergence decision. Full OAuth2/OIDC per §4 remains an owner-scoped stage of its own plus the §9 auth ADR. |
 | Security-Header, Rate Limiting — §4 | **met at the edge** (2026-08-05, G7) for the modern stand · never for `legacy/` | Delivered by the Traefik overlay ([ADR-0014](adr/0014-authentifizierung-am-edge.md)), verified by `modern/edge/verify-edge.sh`, not asserted: `X-Frame-Options: DENY`, `X-Content-Type-Options: nosniff`, `Referrer-Policy: no-referrer`, a CSP whose `script-src` is strict, and a token-bucket rate limit measured firing (200 concurrent requests → 103–173 × 429 across five runs; the same burst straight at the application → 0). HSTS ships **off** (`stsSeconds: 0`) on purpose: enabled over plain HTTP it teaches the browser to refuse `http://localhost` for a year; the TLS-terminating host sets it. `legacy/` gets none of this — it is the exhibit. |
 | CSRF-Schutz — §4 | **deliberately not implemented**, reason recorded | Not an oversight and not deferred-by-drift: the application has no session, no cookie and no ambient authority, so a classic CSRF token would protect nothing. What stage 6 *did* introduce is Basic auth at the edge, and browsers cache those credentials — a cross-site POST could ride them. That residual risk is written down in `SECURITY.md` and accepted for a demonstration stand; the real fix (session-based OIDC plus token) belongs to the owner-scoped auth stage below, not to an ops stage. Enabling Spring Security's CSRF today has a measured price: 17 characterization write calls plus every E2E write scenario turn red, i.e. it is a pinned-contract change under the ADR-0004 gate. |

modern/pom.xml

@@ -33,6 +33,15 @@
 		     Never raise this above a measured value: a gate nobody can reach is the reason
 		     -DskipTests exists. See docs/DEVIATIONS.md. -->
 		<jacoco.line.minimum>0.80</jacoco.line.minimum>
+		<!-- Security overrides of Boot 4.1.1's managed versions (2026-10-21), each held at the
+		     minimum version that fixes what the image scan reported, not at the newest:
+		     - Tomcat 11.0.24 → 11.0.25: CVE-2026-65182, CVE-2026-65905, CVE-2026-68525 (CRITICAL)
+		     - Jackson 3 3.1.5 → 3.1.7: CVE-2026-89407, -89425, -68497, -91776, -91777 (HIGH).
+		       3.1.7 stays on Boot's managed minor; 3.2.2 would NOT close -89425/-91776/-91777
+		       (fixed in 3.1.7 or 3.2.3), and a minor jump touches the wire format ADR-0005 pins.
+		     Drop both properties when the Boot parent manages these versions or newer. -->
+		<tomcat.version>11.0.25</tomcat.version>
+		<jackson-bom.version>3.1.7</jackson-bom.version>
 	</properties>
 
 	<dependencyManagement>
@@ -65,7 +74,8 @@
 		     fixed in 42.7.12. Not theoretical: PostgreSQL 18 authenticates with SCRAM, so
 		     this is the mechanism the modern stand actually uses. Held at the minimum fixed
 		     version rather than the newest (42.7.13) so this is a security fix and nothing
-		     else — Dependabot owns routine drift.
+		     else — Dependabot owns routine drift. 2026-10-21: 42.7.14 (Dependabot #101),
+		     testbed pom moved in the same commit.
 		     Declared as a literal here AND in ai-testgen/testbed/modern/pom.xml rather than
 		     as a property: PomDriftGuardTest compares DECLARED coordinates, and a property
 		     override does not travel through the testbed's imported BOM the way it travels
@@ -74,7 +84,7 @@
 		<dependency>
 			<groupId>org.postgresql</groupId>
 			<artifactId>postgresql</artifactId>
-			<version>42.7.12</version>
+			<version>42.7.14</version>
 		</dependency>
 
 		<!-- G7 (stage 6): schema management moves out of the hand-run SQL files