mirror of
https://github.com/azaion/satellite-provider.git
synced 2026-06-21 18:51:13 +00:00
5214a4a647
Step 14 (Security Audit) for cycle 2 — delta scan against the cycle-1 baseline. Verdict remains PASS_WITH_WARNINGS; no Critical/High. Scope: JWT auth boundary (AZ-487) and UAV multipart upload + ImageSharp decode of attacker-controlled bytes (AZ-488). Both new packages (JwtBearer 8.0.21, ImageSharp 3.1.11 in Services.TileDownloader) checked. Cycle-2 delta: * 0 Critical / 0 High * 2 Medium: F-AUTH-2 (iss/aud not validated — by design until admin team publishes values, AZ-487 § Constraints), F-UAV-1 (ImageSharp decode now runs on attacker-controlled bytes — mitigations sufficient; pin to GHSA subscribe-and-bump policy). * 4 Low: F-AUTH-1 (DEV-ONLY secret in appsettings.Development.json — accepted), F-AUTH-3 (rate-limit gap extends to 401 floods — folds into cycle-1 I3), F-UAV-2 (JsonDocument.Parse on signature-validated claims — bounded by Kestrel header cap), D3 (JwtBearer shares D1 patch line). * 1 Informational: F-UAV-3 (reject reasons disclose gate structure — accepted UX trade-off; documented in contract). OWASP refresh: A01 / A07 move from N/A (with caveat) to PASS_WITH_WARNINGS (per-tenant authz absent; iss/aud + revocation gaps tracked). Pre-deploy operational gate added: deploy pipeline must verify JWT_SECRET != DEV-ONLY placeholder before promoting api. Artifacts: dependency_scan.md, static_analysis.md, owasp_review.md, infrastructure_review.md, security_report.md — all appended with a "Cycle 2 Delta" section preserving cycle-1 finding IDs. Co-authored-by: Cursor <cursoragent@cursor.com>
141 lines
17 KiB
Markdown
141 lines
17 KiB
Markdown
# Phase 2 — Static Analysis (SAST)
|
||
|
||
**Date**: 2026-05-11
|
||
**Scope**: All `*.cs` files in production projects (Api, Common, DataAccess, Services.*) plus Tests for false-positive triage. Configuration files (`appsettings*.json`, `docker-compose*.yml`, `Dockerfile`, `.env`).
|
||
**Method**: Pattern-based grep + targeted file review.
|
||
|
||
## Patterns checked
|
||
|
||
| Category | Pattern(s) | Verdict |
|
||
|----------|-----------|---------|
|
||
| SQL injection | `$"SELECT…"`, `+ "WHERE"`, raw `CommandText`, manual SQL string assembly | **Clean** |
|
||
| Command/process injection | `Process.Start`, `ProcessStartInfo`, `cmd.exe`, `/bin/sh`, `UseShellExecute`, `eval`-equivalent | **Clean** |
|
||
| XSS | unsanitized user input flowed to HTML or `Response.Write` | **N/A** — JSON-only API, no HTML rendering |
|
||
| Template injection | Razor / scriban / handlebars on user input | **N/A** — none used |
|
||
| Hardcoded credentials | `password = "…"`, `secret = "…"`, `token = "…"`, `apikey = "…"` in source | See findings S1, S2 |
|
||
| Weak crypto | MD5/SHA1 for passwords, `RNGCryptoServiceProvider` (deprecated), hardcoded keys | **N/A** — no password storage, no crypto code in app |
|
||
| Insecure deserialization | `BinaryFormatter`, `pickle`, untrusted JSON with type-name handling | **Clean** — `System.Text.Json` with default settings; `Newtonsoft.Json` 13.0.4 used only for outbound serialization to Google session-creation endpoint (line `GoogleMapsDownloaderV2.cs`), no deserialization of untrusted inbound JSON |
|
||
| Path traversal | user input flowed into `File.Open`, `Path.Combine` | **Clean** — file paths are computed server-side from validated tile coordinates; no user-supplied path component reaches the filesystem |
|
||
| Sensitive data in logs | passwords, API keys, tokens, PII in log statements | **Clean** — `GlobalExceptionHandler.cs` logs only `Method`, `Path`, `correlationId`; client gets a generic 500 + correlationId. `CorsConfigurationValidator` warning (`PermissiveDefaultWarning`) does not include secrets. There is a deliberate test fixture `GlobalExceptionHandlerTests.cs:23` that uses `"Connection string Host=secret-db;Password=hunter2 failed at line 42"` to verify the handler does NOT echo exception messages back — this is a positive control, not a finding |
|
||
| Verbose error responses | stack traces or internal details returned to clients | **Clean** — `GlobalExceptionHandler` returns RFC 7807 ProblemDetails with `Detail = "An unexpected error occurred. Use the correlationId to look up the server log entry."` |
|
||
| Input validation | numeric ranges, geo coordinates, enum-like strings | See finding S3 |
|
||
| Hardcoded credentials (cycle 2 delta) | `Jwt:Secret` value in `appsettings*.json` | `appsettings.Development.json` ships a clearly-tagged DEV-ONLY placeholder; `appsettings.json` ships `""`. `JWT_SECRET` env-var overrides both. See cycle-2 finding F-AUTH-1. |
|
||
| Authentication / authorization (cycle 2 delta) | endpoint-level Authorize, custom requirement handlers, claim parsing | `Program.cs` applies `.RequireAuthorization()` on every existing endpoint and the GPS-permission policy on the new `/api/satellite/upload`. `PermissionsAuthorizationHandler` uses `string.Equals(..., Ordinal)` — no substring / case-confusion bypass. See cycle-2 findings F-AUTH-2 .. F-AUTH-4. |
|
||
| Multipart binary input (cycle 2 delta) | uploaded bytes flowing into image decode / file write | `UavTileQualityGate` runs magic-byte check before ImageSharp, wraps decode in scoped `try/catch` for `UnknownImageFormatException` / `InvalidImageContentException`. File path is built from integer coords only via `UavTileUploadHandler.BuildUavTileFilePath`. See cycle-2 finding F-UAV-1. |
|
||
| Untrusted JSON via claims (cycle 2 delta) | `JsonDocument.Parse(claim.Value)` in `PermissionsAuthorizationHandler` | Tokens are signature-validated *before* the handler runs, so the JSON parsed here is already framework-validated bytes from a verified token. Token size is bounded by Kestrel header limits. See cycle-2 finding F-UAV-2. |
|
||
|
||
## Findings
|
||
|
||
### S1 — Default DB password committed in `appsettings.json` (Medium)
|
||
|
||
- **Location**: `SatelliteProvider.Api/appsettings.json:24`
|
||
- **Vulnerable code**:
|
||
```json
|
||
"DefaultConnection": "Host=localhost;Database=satelliteprovider;Username=postgres;Password=postgres"
|
||
```
|
||
- **Description**: The default (non-Development) appsettings file ships with a weak, well-known password (`postgres/postgres`). In production this string is overridden by `ConnectionStrings__DefaultConnection` in `docker-compose.yml`/env, but the file itself becomes the fallback if env-var injection ever fails or is misconfigured (silent connect-as-default behaviour).
|
||
- **Impact**: If a deployment misconfiguration drops the env override, the app silently falls back to attempting `postgres:postgres@localhost`. On a developer workstation this connects to the local Postgres container with full superuser; in production it would fail loudly only if the prod DB has different creds. Combined with finding S2 below (matching weak creds in compose file), this normalises a credential pattern that real production deployments may inherit.
|
||
- **Remediation**:
|
||
- Replace the default value with a deliberately-invalid placeholder such as `Host=__set-via-env__;Database=__;Username=__;Password=__` so a misconfiguration fails fast at startup instead of silently falling through.
|
||
- OR remove the `ConnectionStrings:DefaultConnection` key from `appsettings.json` entirely and require the env var; `Program.cs` line 23–24 already throws when missing — keep that behaviour.
|
||
|
||
### S2 — Weak Postgres credentials in `docker-compose.yml` (Medium, dev-only as written)
|
||
|
||
- **Location**: `docker-compose.yml:6-7, 30`
|
||
- **Vulnerable code**:
|
||
```yaml
|
||
POSTGRES_USER: postgres
|
||
POSTGRES_PASSWORD: postgres
|
||
…
|
||
- ConnectionStrings__DefaultConnection=Host=postgres;Port=5432;Database=satelliteprovider;Username=postgres;Password=postgres
|
||
```
|
||
- **Description**: Same `postgres/postgres` credentials as S1. The compose file is labelled `Development` (`ASPNETCORE_ENVIRONMENT=Development`), so this is contained — but the file is the only compose artifact in the repo, which means anyone running `docker-compose up` on a network-reachable host immediately exposes a Postgres-with-default-creds.
|
||
- **Impact**: Postgres on `0.0.0.0:5432` (port `"5432:5432"` mapping) with `postgres/postgres` is one of the most-scanned credential pairs on the public internet. If a developer runs this on a non-laptop host (cloud VM, shared lab, etc.) the DB is trivially compromised within minutes.
|
||
- **Remediation**:
|
||
- Bind `5432` to `127.0.0.1:5432` rather than `0.0.0.0:5432` so the host firewall isn't the only protection. (Replace `"5432:5432"` with `"127.0.0.1:5432:5432"`.)
|
||
- Source `POSTGRES_USER` / `POSTGRES_PASSWORD` from the same `.env` file that already supplies `GOOGLE_MAPS_API_KEY` (line 31 already shows the pattern). Provide an `.env.example` with placeholder values and document the required vars in the README.
|
||
- The deploy/observability docs at `_docs/02_document/deployment/` already describe a secret-manager strategy for staging/prod — fold the same pattern into the dev compose.
|
||
|
||
### S3 — Latitude / longitude inputs not range-validated at the API boundary (Low)
|
||
|
||
- **Locations**:
|
||
- `SatelliteProvider.Api/Program.cs:169` — `GetTileByLatLon([FromQuery] double Latitude, [FromQuery] double Longitude, [FromQuery] int ZoomLevel, …)`
|
||
- `SatelliteProvider.Api/Program.cs:207` — `RequestRegion` validates `SizeMeters` only; `request.Latitude` / `request.Longitude` are unchecked
|
||
- `SatelliteProvider.Api/Program.cs:237` — `CreateRoute` delegates to `RouteService` which validates names but does not range-check waypoint coordinates
|
||
- **Description**: `Latitude`, `Longitude`, and (for region requests) the implicit `MaxRoutePointSpacingMeters` boundary are accepted without enforcing valid geographic ranges (`-90 ≤ lat ≤ 90`, `-180 ≤ lon ≤ 180`). `ZoomLevel` IS validated downstream by `GoogleMapsDownloaderV2` against `MapConfig.AllowedZoomLevels` — so it is fine.
|
||
- **Impact**:
|
||
- Garbage inputs (e.g. `lat=999`) propagate through `GeoUtils.WorldToTilePos` and the slippy-map math, eventually producing nonsensical tile coordinates that are persisted to `tiles` and `regions`. This is a **data-quality** issue, not a code-execution issue.
|
||
- No DoS amplification: every tile-download endpoint already enforces zoom against `AllowedZoomLevels`, so an attacker cannot use lat/lon abuse to multiply outbound Google Maps traffic beyond what zoom already bounds.
|
||
- **Remediation**: Add explicit guard clauses at the API boundary (matches the existing `SizeMeters` 100-10000 pattern):
|
||
```csharp
|
||
if (Latitude < -90 || Latitude > 90) return Results.BadRequest(new { error = "Latitude must be between -90 and 90" });
|
||
if (Longitude < -180 || Longitude > 180) return Results.BadRequest(new { error = "Longitude must be between -180 and 180" });
|
||
```
|
||
Apply uniformly to `GetTileByLatLon`, `RequestRegion`, and to each waypoint inside `CreateRoute`.
|
||
|
||
### S4 — `.env` file on developer filesystem contains an apparently real Google Maps API key (Medium — exposure depends on key reach)
|
||
|
||
- **Location**: `.env` (workspace root, **not** tracked — confirmed via `git ls-files` and `.gitignore:10`)
|
||
- **Description**: The local `.env` contains a 39-character `AIzaSy…` value matching the Google Maps API key format. The file is correctly excluded from git (line 10 of `.gitignore`) and `git log -- .env` returns no history, so the key was never committed to this repository.
|
||
- **Impact**: No repository exposure. **However**:
|
||
- If the same key is shared across developers via Slack / email / other repos, it has likely already leaked elsewhere.
|
||
- There is no `.env.example` template in the repo, which means new contributors typically request the real key via insecure channels rather than generating a fresh one.
|
||
- The key has no per-call attribution; abuse cannot be traced back to a specific developer.
|
||
- **Remediation**:
|
||
- **Rotate the key in the Google Cloud console** (out of scope for this audit — the key value is intentionally not echoed into this report).
|
||
- Add `.env.example` to the repo with `GOOGLE_MAPS_API_KEY=replace-with-your-own-key-from-cloud-console` and reference it in the README setup section.
|
||
- Configure Google Cloud key restrictions: HTTP referrer allowlist (for browser keys) or IP allowlist (for server keys), and per-API quotas. Optional: per-developer keys.
|
||
|
||
---
|
||
|
||
## Cycle 2 Delta Findings (AZ-487 + AZ-488)
|
||
|
||
### F-AUTH-1 — Dev JWT secret is committed to `appsettings.Development.json` (Low — accepted by design)
|
||
|
||
- **Location**: `SatelliteProvider.Api/appsettings.Development.json:14` — `"Secret": "DEV-ONLY-DO-NOT-USE-IN-PROD-replace-with-real-secret-via-JWT_SECRET-env-var"`.
|
||
- **Description**: A 73-byte placeholder labelled DEV-ONLY ships in the repo. The value is clearly tagged; `ResolveSecretOrThrow` in `AuthenticationServiceCollectionExtensions.cs:43` reads `JWT_SECRET` from the environment first and only falls back to config when it is unset, so a production deploy with `JWT_SECRET` set overrides it.
|
||
- **Impact**: Cosmetic only — the placeholder is not a usable production secret (it is published on every git clone and would be rejected by any token verifier already in the wild). A careless operator who copies the file verbatim into prod and forgets to set `JWT_SECRET` would still pass the ≥32-byte gate, so the secret would *work* locally — that is the dependency to monitor.
|
||
- **Disposition**: Accept. Mitigation: the `DEV-ONLY-DO-NOT-USE-IN-PROD` prefix is the operator-readable warning; the deploy skill must verify `JWT_SECRET` is set before promotion.
|
||
|
||
### F-AUTH-2 — JWT issuer / audience are not validated (Medium — by design, until admin team defines values)
|
||
|
||
- **Location**: `SatelliteProvider.Api/Authentication/AuthenticationServiceCollectionExtensions.cs:31-32` — `ValidateIssuer = false`, `ValidateAudience = false`.
|
||
- **Description**: Per the suite contract `suite/_docs/10_auth.md`, expected `iss` / `aud` values are not yet defined. The validator therefore accepts any HS256 token signed with the correct shared secret — including tokens minted by other services in the suite that share the secret. This is a horizontal-trust risk: any service that holds `JWT_SECRET` can mint tokens accepted by satellite-provider as if they came from the admin API.
|
||
- **Impact**: Bounded by the secret-distribution policy. Within the trust boundary documented in cycle 1's A01 caveat ("internal/trusted-network service") this is acceptable.
|
||
- **Remediation (follow-up, NOT this cycle)**: When the admin team publishes `iss` / `aud` values, flip `ValidateIssuer = true` + `ValidIssuer = "<admin-iss>"` and the audience equivalent in `AddSatelliteJwt`. AZ-487 § Constraints already flags this as a small follow-up.
|
||
|
||
### F-AUTH-3 — No rate limiting on 401-producing paths (Low — recurrence of cycle-1 I3)
|
||
|
||
- **Location**: every `/api/satellite/*` endpoint after the AZ-487 `.RequireAuthorization()` middleware.
|
||
- **Description**: An attacker can flood `Authorization: Bearer <random>` requests; each one triggers an HMAC verification (cheap, but non-zero) and an HTTP 401 response. This re-uses the cycle-1 I3 finding ("no inbound rate limiting on any HTTP endpoint") — the JWT layer didn't introduce a new vulnerability, but it did add a new cheap-to-trigger 401 surface that magnifies I3.
|
||
- **Disposition**: Track under existing I3 remediation (wire `Microsoft.AspNetCore.RateLimiting`). No separate Jira.
|
||
|
||
### F-UAV-1 — ImageSharp decode on attacker-controlled bytes (Medium — exposure increase, mitigations sufficient today)
|
||
|
||
- **Location**: `SatelliteProvider.Services.TileDownloader/UavTileQualityGate.cs:60-95` — `Image.Identify` (Rule 3) and `Image.Load<L8>` + `Mutate(ctx => ctx.Resize)` (Rule 5).
|
||
- **Description**: Pre-AZ-488, ImageSharp only decoded responses from the Google Maps tile CDN (trusted origin). AZ-488 added a second call site that decodes arbitrary `POST /api/satellite/upload` payloads. Current ImageSharp 3.1.11 is patched (see cycle-2 dependency-scan finding F-DEPS-UAV); the change here is *exposure*, not a present vulnerability.
|
||
- **Mitigations in place**:
|
||
- Rule 1 magic-byte gate runs before any ImageSharp call (`FF D8 FF` prefix required).
|
||
- Rule 2 caps per-item size at 5 MiB; Kestrel + FormOptions cap the envelope at `MaxBatchSize × MaxBytes`.
|
||
- Decode is wrapped in `try { … } catch (UnknownImageFormatException) { … } catch (InvalidImageContentException) { … }` — malformed JPEGs produce a structured `INVALID_FORMAT` reject; no unhandled exception reaches the client.
|
||
- **Remediation**: Subscribe to `SixLabors.ImageSharp` GHSA advisories; bump within 7 days of a patch. Sandboxing (separate process / libvips + seccomp) is not warranted at the current trust boundary but should be reconsidered if the endpoint is exposed publicly. Recorded as recurring follow-up.
|
||
|
||
### F-UAV-2 — `JsonDocument.Parse` invoked on token-supplied claim values (Low — bounded by Kestrel header limits)
|
||
|
||
- **Location**: `SatelliteProvider.Api/Authentication/PermissionsRequirement.cs:84-111` — `JsonDocument.Parse(claim.Value)` when the `permissions` claim arrives as a JSON-array string.
|
||
- **Description**: `JsonDocument.Parse` has no built-in depth or size limit. A maliciously-shaped permissions claim (e.g. deeply-nested array) would consume CPU/heap during parsing. The token has already passed HS256 signature validation by the time the handler runs, so this is only exploitable by a party that holds `JWT_SECRET` — i.e. another suite service or an admin-team principal — and only inside the issued-token-size window (bounded by Kestrel's `MaxRequestHeadersTotalSize`, default 32 KiB).
|
||
- **Disposition**: Accept. The combination of `RequireSignedTokens = true` + header-size cap + ordinal-only string comparison makes a practical exploit prohibitive. Future hardening: pass `JsonDocumentOptions { MaxDepth = 8 }` to `JsonDocument.Parse` and reject claims longer than e.g. 8 KiB before parsing.
|
||
|
||
### F-UAV-3 — Reject reasons disclose gate structure (Informational — accepted trade-off)
|
||
|
||
- **Location**: `SatelliteProvider.Services.TileDownloader/UavTileQualityGate.cs` — each rule returns a distinct enum code.
|
||
- **Description**: A client (or attacker who can present a `GPS`-permission token) can map the gate by probing inputs (1×1 black image → `WRONG_DIMENSIONS`; 1 KB JPEG → `SIZE_OUT_OF_BAND`; etc.). The thresholds are also documented in the public contract `_docs/02_document/contracts/api/uav-tile-upload.md`.
|
||
- **Disposition**: Accept — UX (helping clients self-correct) outweighs the information-hiding benefit, especially since the contract is public anyway. Flagged to keep operators aware: rule thresholds are NOT a security boundary; do not move secrets into reject details.
|
||
|
||
## Self-verification
|
||
|
||
- [x] All production source directories scanned (Api, Common, DataAccess, Services.TileDownloader, Services.RegionProcessing, Services.RouteManagement)
|
||
- [x] Each finding has file path and line number
|
||
- [x] False positives from test files explicitly distinguished (`GlobalExceptionHandlerTests.cs:23` "leakySecret" is a positive control)
|
||
- [x] No real secret values printed in this report (S4 is described without echoing the key)
|