Close the read-side privilege escalation and fix proxy-aware IPs (0.44.0)

F1 — Any authenticated user could read any file the backend could see.
/api/files/read and /download hung on get_current_user, and the sandbox that
should have caught that was open by default: ALLOWED_BROWSE_ROOTS contained
"/", for which _is_allowed() waves through every path. So the `user` role could
download stackpilot.db (password hashes, agent tokens, backup credentials),
every stack's .env and every .secrets/* file — with no audit trail, because
only mutations were logged.

Implementing that turned up three more doors into the same room, all fixed
here since closing only the first would have made the fix cosmetic:
GET /api/stacks/{id} handed the .env to any user, /export tarred the whole
stack dir including .secrets/*, and both the agent file proxies and
/api/agents/{id}/stacks/{id} repeated the leak for every remote host. All 24
filesystem-touching routes are now admin-only; reads and downloads are audited
(listing is not — the Files page polls it). DATA_DIR is refused outright, since
the API deliberately masks agent tokens and destination secrets and the browser
would otherwise be the way around that. "/" is out of the default browse roots.

F2 — Backup destination credentials were plaintext JSON in the DB, which is
what made F1 worth exploiting. They are now Fernet-encrypted at rest behind
parse_config/dump_config, with existing rows migrated at startup.

This needed a prerequisite from F6: the key is derived from SECRET_KEY, which
was regenerated on every boot when unset. Encrypting against a key that changes
per restart would be worse than plaintext, so an auto-generated SECRET_KEY is
now persisted to ${DATA_DIR}/secret_key at mode 0600. Sessions surviving a
restart is a welcome side effect.

F3 — /api/audit is admin-only. Also hidden from the dashboard and the nav for
non-admins, so nobody polls into a 403.

F4 — uvicorn now runs with --proxy-headers, so nginx's X-Forwarded-For is
honoured. Without it request.client.host was the frontend container's IP for
every request, which made the login rate limit global instead of per-IP (10
failures locked out everyone) and filled the audit log's IP column with one
useless value.

Verified: encrypt/decrypt round-trip incl. plaintext passthrough, idempotent
re-encryption and wrong-key handling; sandbox denial for DATA_DIR, traversal
into it, and paths outside the roots, with the allowed roots still reachable.
Both against stubbed settings — there is no Docker here, so nothing was run
end to end.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dk43rmEeRfYi5wsLDbmfyG

backend/Dockerfile

@@ -27,4 +27,15 @@ EXPOSE 5008
 HEALTHCHECK --interval=30s --timeout=5s --start-period=10s \
     CMD curl -fsS http://localhost:5008/api/health || exit 1
 
-CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "5008"]
+# --proxy-headers: nginx setzt X-Forwarded-For, ohne dieses Flag ignoriert
+# uvicorn den Header und request.client.host ist fuer JEDE Anfrage die IP des
+# Frontend-Containers -- was das Login-Rate-Limit global statt pro IP wirken
+# laesst und die IP-Spalte im Audit-Log wertlos macht.
+#
+# forwarded-allow-ips=* vertraut dem Header von jedem Absender. Das ist hier
+# richtig, weil der Backend-Port nur im Docker-Netz erreichbar ist (siehe
+# "expose" statt "ports" in docker-compose.yml). Wer 5008 direkt nach aussen
+# gibt, muss den Wert auf die IP des eigenen Proxys einschraenken -- sonst
+# kann ein Client seine eigene Herkunfts-IP faelschen.
+CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "5008", \
+     "--proxy-headers", "--forwarded-allow-ips", "*"]

backend/config.py

@@ -1,11 +1,13 @@
 """Application settings, loaded from environment variables."""
 from __future__ import annotations
 
+import os
 import secrets
+import stat
 from functools import lru_cache
 from typing import Annotated
 
-from pydantic import field_validator
+from pydantic import ValidationInfo, field_validator
 from pydantic_settings import BaseSettings, NoDecode, SettingsConfigDict
 
 
@@ -17,7 +19,8 @@ class Settings(BaseSettings):
     DATA_DIR: str = "/opt/stackpilot/data"
 
     # Security
-    SECRET_KEY: str = ""  # Auto-generated if empty (dev only); set in prod.
+    # Auto-generated and persisted to ${DATA_DIR}/secret_key when left empty.
+    SECRET_KEY: str = ""
     ALGORITHM: str = "HS256"
     ACCESS_TOKEN_EXPIRE_MINUTES: int = 60
     REFRESH_TOKEN_EXPIRE_DAYS: int = 30
@@ -39,9 +42,11 @@ class Settings(BaseSettings):
     # Only used when running the agent app (agent_app:app).
     AGENT_TOKEN: str = ""
 
-    # Host browser sandbox roots
+    # Host browser sandbox roots. Deliberately does NOT contain "/": that entry
+    # makes _is_allowed() wave through every path, i.e. it switches the sandbox
+    # off. Add it back explicitly if you really want the whole filesystem.
     ALLOWED_BROWSE_ROOTS: Annotated[list[str], NoDecode] = [
-        "/", "/mnt", "/media", "/srv", "/opt",
+        "/mnt", "/media", "/srv", "/opt", "/home",
     ]
     HOST_ROOT_PREFIX: str = ""  # e.g. "/host_root" when host / is bind-mounted
 
@@ -55,8 +60,37 @@ class Settings(BaseSettings):
 
     @field_validator("SECRET_KEY", mode="after")
     @classmethod
-    def _ensure_secret(cls, v: str) -> str:
-        return v or secrets.token_urlsafe(48)
+    def _ensure_secret(cls, v: str, info: ValidationInfo) -> str:
+        """Return the configured key, or a persisted auto-generated one.
+
+        Generating a fresh key per process (the old behaviour) silently
+        invalidated every session on each restart, and would now also make the
+        encrypted backup-destination credentials undecryptable. So the
+        generated key is written next to the database instead, mode 0600, and
+        read back on the next start. An explicitly configured SECRET_KEY always
+        wins and nothing is written.
+        """
+        if v:
+            return v
+        data_dir = info.data.get("DATA_DIR") or "/opt/stackpilot/data"
+        key_file = os.path.join(data_dir, "secret_key")
+        try:
+            with open(key_file, "r", encoding="utf-8") as fh:
+                if existing := fh.read().strip():
+                    return existing
+        except OSError:
+            pass
+        generated = secrets.token_urlsafe(48)
+        try:
+            os.makedirs(data_dir, exist_ok=True)
+            with open(key_file, "w", encoding="utf-8") as fh:
+                fh.write(generated + "\n")
+            os.chmod(key_file, stat.S_IRUSR | stat.S_IWUSR)
+        except OSError:
+            # Read-only data dir: fall back to the old per-process behaviour
+            # rather than refusing to boot. Sessions won't survive a restart.
+            pass
+        return generated
 
     @field_validator(
         "NOTIFY_WEBHOOKS", "ALLOWED_BROWSE_ROOTS", "CORS_ORIGINS", mode="before"

backend/main.py

@@ -36,7 +36,12 @@ from routers import (
     volumes,
     ws,
 )
-from services import schedule_service, template_service, update_service
+from services import (
+    backup_destination_service,
+    schedule_service,
+    template_service,
+    update_service,
+)
 
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger("stackpilot")
@@ -51,6 +56,15 @@ async def lifespan(app: FastAPI):
             stacks.sync_discovered_stacks(session)
     except Exception as exc:  # noqa: BLE001
         logger.warning("Stack discovery failed: %s", exc)
+    # One-off: encrypt backup-destination credentials written before they were
+    # stored encrypted (see services/crypto_service.py).
+    try:
+        with Session(engine) as session:
+            encrypted = backup_destination_service.migrate_plaintext_configs(session)
+        if encrypted:
+            logger.info("Encrypted %d backup destination config(s) at rest", encrypted)
+    except Exception as exc:  # noqa: BLE001
+        logger.warning("Destination config encryption migration failed: %s", exc)
     try:
         moved = template_service.migrate_legacy_db_templates()
         if moved:

backend/requirements.txt

@@ -5,6 +5,8 @@ sqlmodel==0.0.22
 pydantic==2.10.4
 pydantic-settings==2.7.1
 python-jose[cryptography]==3.3.0
+# Direct dependency: services/crypto_service encrypts DB-stored secrets.
+cryptography==44.0.0
 passlib[bcrypt]==1.7.4
 bcrypt==4.2.1
 python-multipart==0.0.20

backend/routers/agents.py

@@ -245,10 +245,14 @@ async def agent_stack_detail(
     agent_id: int,
     stack_id: str,
     session: Session = Depends(get_session),
-    _user: User = Depends(get_current_user),
+    user: User = Depends(get_current_user),
 ) -> dict:
     agent = _get_or_404(session, agent_id)
     data = await _proxy(session, agent, "GET", f"/agent/stacks/{stack_id}")
+    # Withhold the .env from the read-only role, exactly as the local
+    # GET /api/stacks/{id} does.
+    if user.role != "admin" and isinstance(data, dict):
+        data["env"] = ""
     data["agent_id"] = agent.id
     data["agent_name"] = agent.name
     return data
@@ -796,7 +800,7 @@ async def agent_files_list(
     path: str = Query("/"),
     show_hidden: bool = Query(False),
     session: Session = Depends(get_session),
-    _user: User = Depends(get_current_user),
+    _admin: User = Depends(require_admin),
 ) -> dict:
     agent = _get_or_404(session, agent_id)
     return await _proxy(
@@ -808,22 +812,33 @@ async def agent_files_list(
 @router.get("/{agent_id}/files/read")
 async def agent_files_read(
     agent_id: int,
+    request: Request,
     path: str = Query(...),
     session: Session = Depends(get_session),
-    _user: User = Depends(get_current_user),
+    user: User = Depends(require_admin),
 ) -> dict:
     agent = _get_or_404(session, agent_id)
-    return await _proxy(session, agent, "GET", "/agent/files/read", params={"path": path})
+    result = await _proxy(session, agent, "GET", "/agent/files/read", params={"path": path})
+    audit_service.record(
+        session, user=user.username, action="file.read",
+        target=f"{agent.name}:{path}", ip=_ip(request),
+    )
+    return result
 
 
 @router.get("/{agent_id}/files/download")
 async def agent_files_download(
     agent_id: int,
+    request: Request,
     path: str = Query(...),
     session: Session = Depends(get_session),
-    _user: User = Depends(get_current_user),
+    user: User = Depends(require_admin),
 ):
     agent = _get_or_404(session, agent_id)
+    audit_service.record(
+        session, user=user.username, action="file.download",
+        target=f"{agent.name}:{path}", ip=_ip(request),
+    )
     # Stream the agent's response straight through (works for single files and
     # for on-the-fly folder zips), so nothing is staged to disk and the
     # download starts immediately. Pull the first chunk eagerly so a failed

backend/routers/audit.py

@@ -1,4 +1,9 @@
-"""Audit log query endpoint."""
+"""Audit log query endpoint.
+
+Admin-only: the log is security telemetry (who did what, from which IP,
+including every administrator's activity) and has no business being readable
+by an account with the ``user`` role.
+"""
 from __future__ import annotations
 
 from typing import Optional
@@ -6,7 +11,7 @@ from typing import Optional
 from fastapi import APIRouter, Depends, Query
 from sqlmodel import Session, select
 
-from auth import get_current_user
+from auth import require_admin
 from database import get_session
 from models.audit import AuditLog
 from models.user import User
@@ -20,7 +25,7 @@ def list_audit(
     offset: int = 0,
     stack_id: Optional[str] = None,
     session: Session = Depends(get_session),
-    _user: User = Depends(get_current_user),
+    _admin: User = Depends(require_admin),
 ) -> list[AuditLog]:
     stmt = select(AuditLog).order_by(AuditLog.timestamp.desc())
     if stack_id:

backend/routers/destinations.py

@@ -2,12 +2,11 @@
 from __future__ import annotations
 
 import asyncio
-import json
 
 from fastapi import APIRouter, Depends, HTTPException, Request
 from sqlmodel import Session, select
 
-from auth import get_current_user, require_admin
+from auth import require_admin
 from database import get_session
 from models.backup_destination import (
     DESTINATION_TYPES,
@@ -66,7 +65,9 @@ def create_destination(
 ) -> DestinationRead:
     if body.type not in DESTINATION_TYPES:
         raise HTTPException(status_code=400, detail=f"Unknown type '{body.type}'")
-    d = BackupDestination(name=body.name, type=body.type, config=json.dumps(body.config))
+    d = BackupDestination(
+        name=body.name, type=body.type, config=dest_service.dump_config(body.config)
+    )
     session.add(d)
     session.commit()
     session.refresh(d)
@@ -95,7 +96,7 @@ def update_destination(
             if k in SECRET_KEYS and (v == "" or v == "••••••"):
                 continue  # keep existing secret
             existing[k] = v
-        d.config = json.dumps(existing)
+        d.config = dest_service.dump_config(existing)
     session.add(d)
     session.commit()
     session.refresh(d)

backend/routers/files.py

@@ -1,7 +1,10 @@
 """Full host filesystem browser: list, read, edit, manage, up/download.
 
-Listing and reads require an authenticated user; every mutating operation
-(write, mkdir, rename, delete, upload) requires admin and is audit-logged.
+Every operation requires admin. Reads are not less dangerous than writes here:
+the browser reaches whatever the backend container can see, which includes
+every stack's ``.env`` and ``.secrets/*``. Reading a file and downloading one
+are audit-logged just like the mutating operations; directory listing is not,
+because the Files page polls it and would drown the log.
 All paths are sandboxed by :mod:`services.file_service`.
 """
 from __future__ import annotations
@@ -23,7 +26,7 @@ from fastapi.responses import FileResponse, StreamingResponse
 from pydantic import BaseModel
 from sqlmodel import Session
 
-from auth import get_current_user, require_admin
+from auth import require_admin
 from database import get_session
 from models.user import User
 from services import audit_service, device_service, file_service
@@ -57,24 +60,35 @@ def _guard(fn, *args, **kwargs):
 def list_dir(
     path: str = Query("/"),
     show_hidden: bool = Query(False),
-    _user: User = Depends(get_current_user),
+    _admin: User = Depends(require_admin),
 ) -> dict:
     return _guard(device_service.browse, path, show_hidden)
 
 
 @router.get("/read")
 def read_file(
+    request: Request,
     path: str = Query(...),
-    _user: User = Depends(get_current_user),
+    session: Session = Depends(get_session),
+    user: User = Depends(require_admin),
 ) -> dict:
-    return _guard(file_service.read_file, path)
+    result = _guard(file_service.read_file, path)
+    audit_service.record(
+        session, user=user.username, action="file.read", target=path, ip=_ip(request)
+    )
+    return result
 
 
 @router.get("/download")
 def download(
+    request: Request,
     path: str = Query(...),
-    _user: User = Depends(get_current_user),
+    session: Session = Depends(get_session),
+    user: User = Depends(require_admin),
 ):
+    audit_service.record(
+        session, user=user.username, action="file.download", target=path, ip=_ip(request)
+    )
     if _guard(file_service.is_dir, path):
         filename, chunks = _guard(file_service.open_archive, path)
         # Stream the zip as it's built so the response starts immediately

backend/routers/stacks.py

@@ -155,7 +155,7 @@ def stacks_updates(_user: User = Depends(get_current_user)) -> dict:
 def get_stack(
     stack_id: str,
     session: Session = Depends(get_session),
-    _user: User = Depends(get_current_user),
+    user: User = Depends(get_current_user),
 ) -> dict:
     stack = _get_stack_or_404(session, stack_id)
     try:
@@ -171,7 +171,10 @@ def get_stack(
         "description": stack.description,
         "status": status,
         "yaml": compose_service.read_compose(stack_id),
-        "env": compose_service.read_env(stack_id),
+        # The .env is where credentials live by convention, so it is withheld
+        # from the read-only role — same reasoning as the admin-only file
+        # browser. Non-admins still get status, services and the compose file.
+        "env": compose_service.read_env(stack_id) if user.role == "admin" else "",
         "containers": containers,
         "created_at": stack.created_at,
         "updated_at": stack.updated_at,
@@ -380,8 +383,10 @@ async def service_logs(
 def export_stack(
     stack_id: str,
     session: Session = Depends(get_session),
-    _user: User = Depends(get_current_user),
+    _admin: User = Depends(require_admin),
 ):
+    """Download the whole stack folder as a tarball. Admin only: the archive
+    contains the ``.env`` and every ``.secrets/*`` file verbatim."""
     import io
     import tarfile
     import tempfile

backend/routers/volumes.py

@@ -98,8 +98,11 @@ def generate_yaml(
 def host_paths(
     path: str = Query("/"),
     show_hidden: bool = Query(False),
-    _user: User = Depends(get_current_user),
+    _admin: User = Depends(require_admin),
 ) -> dict:
+    """Directory picker for the volume wizard. Same browse() as the file
+    browser, so it carries the same admin requirement — and only admins can
+    create a volume with the result anyway."""
     try:
         return device_service.browse(path, show_hidden)
     except device_service.BrowseError as exc:

backend/services/backup_destination_service.py

@@ -1,8 +1,11 @@
 """Push/pull stack backups to remote destinations (SFTP, S3-compatible or NFS).
 
 All operations are synchronous (paramiko / boto3 / docker); async callers
-should wrap them with ``asyncio.to_thread``. Destination config is a plain
-dict parsed from the ``BackupDestination.config`` JSON column.
+should wrap them with ``asyncio.to_thread``. Destination config is a dict
+stored in the ``BackupDestination.config`` column as JSON, encrypted at rest
+(:mod:`services.crypto_service`) because it carries SFTP passwords, SSH keys
+and S3 secret keys. Always go through :func:`parse_config` / :func:`dump_config`
+— never touch the column directly.
 """
 from __future__ import annotations
 
@@ -17,6 +20,7 @@ import tempfile
 from typing import Any
 
 from models.backup_destination import BackupDestination
+from services import crypto_service
 
 logger = logging.getLogger("stackpilot.backup_dest")
 
@@ -26,12 +30,48 @@ class DestinationError(Exception):
 
 
 def parse_config(dest: BackupDestination) -> dict:
+    """Decrypt and parse a destination's config.
+
+    Tolerates plaintext (pre-encryption rows) and returns ``{}`` rather than
+    raising if the value can't be decrypted — a destination whose key is gone
+    should show up as unconfigured in the UI, not take the whole list down with
+    a 500. The failure is logged with the destination name so it's findable.
+    """
     try:
-        return json.loads(dest.config or "{}")
+        raw = crypto_service.decrypt(dest.config or "{}")
+    except crypto_service.DecryptError as exc:
+        logger.error("Destination '%s': %s", dest.name, exc)
+        return {}
+    try:
+        return json.loads(raw or "{}")
     except json.JSONDecodeError:
         return {}
 
 
+def dump_config(config: dict) -> str:
+    """Serialise and encrypt a config dict for storage."""
+    return crypto_service.encrypt(json.dumps(config or {}))
+
+
+def migrate_plaintext_configs(session) -> int:
+    """Encrypt destination configs written before encryption existed.
+
+    Runs once at startup. Returns how many rows were rewritten.
+    """
+    from sqlmodel import select
+
+    migrated = 0
+    for dest in session.exec(select(BackupDestination)).all():
+        if crypto_service.is_encrypted(dest.config):
+            continue
+        dest.config = crypto_service.encrypt(dest.config or "{}")
+        session.add(dest)
+        migrated += 1
+    if migrated:
+        session.commit()
+    return migrated
+
+
 # --------------------------------------------------------------------------- #
 # SFTP (paramiko)
 # --------------------------------------------------------------------------- #

backend/services/crypto_service.py

@@ -0,0 +1,76 @@
+"""Symmetric encryption for secrets that have to live in the database.
+
+Most of StackPilot's secrets are files on disk (``.env``, ``.secrets/*``) where
+filesystem permissions are the right control. A few can't be: backup
+destination credentials and agent tokens are needed by background jobs, so they
+sit in ``stackpilot.db``. This module encrypts those at rest.
+
+The key is derived from ``SECRET_KEY`` rather than being a second thing to
+configure — which is exactly why ``SECRET_KEY`` is now persisted (see
+``config._ensure_secret``): a key that changed on every restart would take the
+ciphertext with it.
+
+Ciphertext is stored with an ``enc:v1:`` prefix so plaintext rows written by
+older versions stay recognisable and can be migrated in place.
+"""
+from __future__ import annotations
+
+import base64
+import hashlib
+import logging
+from typing import Optional
+
+from cryptography.fernet import Fernet, InvalidToken
+
+from config import settings
+
+logger = logging.getLogger("stackpilot.crypto")
+
+PREFIX = "enc:v1:"
+_INFO = b"stackpilot-db-field-encryption-v1"
+
+
+class DecryptError(Exception):
+    """Ciphertext could not be decrypted (usually: SECRET_KEY changed)."""
+
+
+def _fernet() -> Fernet:
+    """Fernet built from a 32-byte key derived from SECRET_KEY.
+
+    Not cached: SECRET_KEY is fixed for the process lifetime, and building a
+    Fernet is a hash plus a base64 encode — cheap enough not to bother.
+    """
+    digest = hashlib.blake2b(
+        settings.SECRET_KEY.encode("utf-8"), key=_INFO, digest_size=32
+    ).digest()
+    return Fernet(base64.urlsafe_b64encode(digest))
+
+
+def is_encrypted(value: Optional[str]) -> bool:
+    return bool(value) and value.startswith(PREFIX)
+
+
+def encrypt(plaintext: str) -> str:
+    """Encrypt a string. Already-encrypted input is returned unchanged."""
+    if is_encrypted(plaintext):
+        return plaintext
+    token = _fernet().encrypt((plaintext or "").encode("utf-8"))
+    return PREFIX + token.decode("ascii")
+
+
+def decrypt(value: str) -> str:
+    """Decrypt a value written by :func:`encrypt`.
+
+    Plaintext (no prefix) is passed straight through, so rows written before
+    encryption existed keep working until the startup migration rewrites them.
+    """
+    if not is_encrypted(value):
+        return value or ""
+    try:
+        return _fernet().decrypt(value[len(PREFIX):].encode("ascii")).decode("utf-8")
+    except (InvalidToken, ValueError) as exc:
+        raise DecryptError(
+            "Could not decrypt a stored secret. This normally means SECRET_KEY "
+            "changed since it was saved — restore the old key, or re-enter the "
+            "affected credentials."
+        ) from exc

backend/services/device_service.py

@@ -87,12 +87,28 @@ def detect_devices() -> dict:
 # --------------------------------------------------------------------------- #
 
 
+class BrowseError(Exception):
+    pass
+
+
 def _real_root(path: str) -> str:
-    """Map a logical host path into the container view (HOST_ROOT_PREFIX)."""
+    """Map a logical host path into the container view (HOST_ROOT_PREFIX).
+
+    Refuses anything that resolves inside StackPilot's own ``DATA_DIR``. That
+    directory holds ``stackpilot.db`` — users, password hashes, agent tokens and
+    backup-destination credentials — and the API deliberately never hands those
+    out (``AgentRead.token_set`` is a bool, destination secrets come back
+    masked). Without this the file browser would be a way around that, for
+    admins too. Note this only bites when ``HOST_ROOT_PREFIX`` is empty: with a
+    prefix set, no logical path can reach the container's own ``/data`` at all.
+    """
     prefix = settings.HOST_ROOT_PREFIX.rstrip("/")
-    if prefix:
-        return prefix + path
-    return path
+    real = prefix + path if prefix else path
+    data_dir = os.path.normpath(settings.DATA_DIR)
+    norm = os.path.normpath(real)
+    if norm == data_dir or norm.startswith(data_dir + os.sep):
+        raise BrowseError("Path is inside StackPilot's own data directory")
+    return real
 
 
 def _is_allowed(path: str) -> bool:
@@ -104,10 +120,6 @@ def _is_allowed(path: str) -> bool:
     return False
 
 
-class BrowseError(Exception):
-    pass
-
-
 def browse(path: str = "/", show_hidden: bool = False) -> dict:
     path = os.path.normpath(path or "/")
     if not path.startswith("/"):

backend/version.py

@@ -1,3 +1,3 @@
 """Single source of truth for the StackPilot release version."""
 
-APP_VERSION = "0.43.0"
+APP_VERSION = "0.44.0"