Security hardening from adversarial review (5-agent workflow)

Fixed confirmed findings before public exposure:
HIGH:
- upload filename path traversal → store.safe_name() strips to basename
- login rate-limit XFF bypass → key on request.client.host + per-username bucket;
  auth.check_login() burns bcrypt time on unknown users (no enumeration)
- cross-user read access → per-user isolation: guests see/use/download/delete only
  their own assets & jobs (owner sees all); WS job events scoped per-user
MEDIUM:
- unbounded upload read → bounded chunked streaming to the 1GB cap
- asset member path check → Path.is_relative_to boundary + ownership gate
- WS token-in-query-string leak → session-cookie-only WS auth
LOW:
- retry_job bypassed the per-user job cap → cap now checked on retry
- wholesale API-key injection → env_for_operator injects a paid key only to
  operators that declare it (guest local jobs never receive fal/OpenRouter keys)
- session revocation → users.session_epoch, bumped on password change
- int() 500s → 400; net-lane defense-in-depth (guests blocked by requires_env AND
  resources==net, so a mis-tagged paid op is still blocked)
+ public /api/health for serve.sh & proxy; docs/VPS.md; mb MB_TOKEN bearer auth

tests/smoke.sh: 34 checks passing incl. all new hardening.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

docs/VPS.md

@@ -0,0 +1,93 @@
+# Public hosting — MODELBEAST at modelbeast.digalot.fyi
+
+Expose MODELBEAST to the internet **without exposing the M3 Ultra directly**: a spare
+VPS is the only public surface and reverse-proxies to the Mac over the tailnet. If the
+VPS is ever compromised you just remove one tailnet node; the Mac is never on the
+public internet.
+
+```
+  friend's browser ──https──▶ VPS (Caddy, public)
+                                  │  reverse_proxy over tailnet
+                                  ▼
+                          M3 Ultra 100.89.131.57:8777  (auth enforced here)
+```
+
+**Prerequisite: auth must be live on the Mac first** (HANDOFF2 Phase 1 — done). Do not
+open this to the internet until `mb`/the UI require a login. Every public request now
+hits the login gate; guests are local-only by hard server-side rule.
+
+## On the VPS (run these — you own that box; I can't reach it)
+
+```bash
+# 1) join the same tailnet as the M3 Ultra
+curl -fsSL https://tailscale.com/install.sh | sh
+sudo tailscale up
+
+# 2) Caddy (automatic HTTPS via Let's Encrypt)
+sudo apt update && sudo apt install -y caddy
+```
+
+Write `/etc/caddy/Caddyfile`:
+```
+modelbeast.digalot.fyi {
+    encode zstd gzip
+    request_body {
+        max_size 1GB          # matches the app's upload cap
+    }
+    reverse_proxy 100.89.131.57:8777
+}
+
+# optional second domain, same backend:
+# modelbeast.partyl.party {
+#     encode zstd gzip
+#     request_body { max_size 1GB }
+#     reverse_proxy 100.89.131.57:8777
+# }
+```
+
+```bash
+sudo systemctl reload caddy
+```
+
+**DNS:** add an `A` record `modelbeast` → the VPS's public IP on `digalot.fyi` (and
+`partyl.party` if using it). Caddy fetches the TLS cert automatically on first request.
+
+## Why these choices
+
+- **Subdomain, not `digalot.fyi/modelbeast`.** A path prefix would force a Vite
+  `base` + FastAPI `root_path` rework across the SPA, the WebSocket, and the `mb` CLI
+  for zero benefit. `modelbeast.<domain>` needs no app changes.
+- **Caddy proxies WebSockets automatically** — the live job-log stream works through
+  it. The session cookie is same-origin (browser talks to the subdomain for both HTTP
+  and WS), so WS auth via cookie just works.
+- **Auth lives in the app, not the proxy.** Proxy basic-auth can't do per-user roles,
+  the guest local-only rule, or per-user job caps. All of that is enforced in FastAPI.
+- **Rejected: Tailscale Funnel.** Zero-VPS and simplest, but no custom domain and it's
+  bandwidth-capped. Keep it in your back pocket as the fallback if the VPS dies:
+  `tailscale funnel 8777` publishes `https://<machine>.<tailnet>.ts.net` instantly.
+
+## Hardening already in the app (verify after going live)
+
+- `Secure` cookie flag is set when the request arrived via https (reads
+  `X-Forwarded-Proto` that Caddy sets) — check DevTools shows `mb_session` as Secure.
+- Upload cap 1 GB in the app **and** in Caddy (`request_body max_size`).
+- Login rate-limit (5/min) keyed on `X-Forwarded-For`.
+- The Mac keeps binding `0.0.0.0` on the tailnet; auth now protects it there too
+  (agents/CLI use `MB_TOKEN`).
+
+## Off-tailnet acceptance test (do this WITH me once Caddy is up)
+
+From a phone on cellular (not on the tailnet):
+1. `https://modelbeast.digalot.fyi` shows the login page over a valid cert.
+2. A guest logs in, runs `flux_local`, watches the live log stream, opens the dashboard.
+3. `MB_HOST=https://modelbeast.digalot.fyi MB_TOKEN=... ./mb ops` works.
+4. Unauthenticated `curl https://modelbeast.digalot.fyi/api/operators` → 401.
+5. A guest gets 403 on `/api/settings` and on a direct cloud-operator job POST.
+
+## Runbook
+- Server stays a single uvicorn worker (the in-process job queue + shared SQLite
+  connection require exactly one worker — do not add `--workers`).
+- Restart the Mac server: `./scripts/serve.sh`. Caddy needs no restart for that.
+- Rotate a leaked session-signing key: delete the `auth_secret` row from the settings
+  table (`sqlite3 data/modelbeast.db "DELETE FROM settings WHERE key='auth_secret'"`)
+  and restart — a new one is generated and all existing sessions are invalidated.

mb

@@ -35,10 +35,17 @@ import uuid
 from pathlib import Path
 
 HOST = os.environ.get("MB_HOST", "http://localhost:8777").rstrip("/")
+TOKEN = os.environ.get("MB_TOKEN", "")
+
+
+def _auth(req):
+    if TOKEN:
+        req.add_header("Authorization", f"Bearer {TOKEN}")
+    return req
 
 
 def http(method, path, body=None):
-    req = urllib.request.Request(HOST + path, method=method)
+    req = _auth(urllib.request.Request(HOST + path, method=method))
     data = None
     if body is not None:
         data = json.dumps(body).encode()
@@ -49,6 +56,9 @@ def http(method, path, body=None):
             raw = r.read()
             return json.loads(raw) if "json" in ct else raw
     except urllib.error.HTTPError as e:
+        if e.code == 401:
+            sys.exit("401 unauthorized — set MB_TOKEN (owner: `mb token create <name>` in the UI, "
+                     "or scripts/users.py). Get one from ⚙ Settings → Users → token.")
         sys.exit(f"HTTP {e.code} {method} {path}: {e.read().decode()[:500]}")
     except urllib.error.URLError as e:
         sys.exit(f"cannot reach {HOST} ({e.reason}) — is the server up? set MB_HOST?")
@@ -63,7 +73,7 @@ def upload_file(path):
     body = (f"--{boundary}\r\nContent-Disposition: form-data; name=\"file\"; "
             f"filename=\"{p.name}\"\r\nContent-Type: {ctype}\r\n\r\n").encode()
     body += p.read_bytes() + f"\r\n--{boundary}--\r\n".encode()
-    req = urllib.request.Request(HOST + "/api/assets", data=body, method="POST")
+    req = _auth(urllib.request.Request(HOST + "/api/assets", data=body, method="POST"))
     req.add_header("Content-Type", f"multipart/form-data; boundary={boundary}")
     try:
         with urllib.request.urlopen(req, timeout=600) as r:
@@ -98,7 +108,7 @@ def download_asset(asset, dest_dir):
     dest_dir = Path(dest_dir)
     dest_dir.mkdir(parents=True, exist_ok=True)
     url = f"{HOST}/api/assets/{asset['id']}/file"
-    with urllib.request.urlopen(url, timeout=600) as r:
+    with urllib.request.urlopen(_auth(urllib.request.Request(url)), timeout=600) as r:
         ct = r.headers.get("content-type", "")
         data = r.read()
     if "json" in ct:
@@ -108,7 +118,7 @@ def download_asset(asset, dest_dir):
             sub.mkdir(exist_ok=True)
             for member in listing["files"]:
                 murl = url + "?member=" + urllib.request.quote(member)
-                with urllib.request.urlopen(murl, timeout=600) as r:
+                with urllib.request.urlopen(_auth(urllib.request.Request(murl)), timeout=600) as r:
                     (sub / member).write_bytes(r.read())
             print(f"  ↓ {sub} ({len(listing['files'])} files)")
             return

scripts/serve.sh

@@ -1,5 +1,9 @@
 #!/usr/bin/env bash
 # Start (or restart) the MODELBEAST server headless on port 8777.
+# NOTE: exactly ONE uvicorn worker (no --workers). The in-process asyncio job
+# queue + the single shared SQLite connection require a single worker; adding
+# workers would fork the queue and corrupt job state. Do not "optimize" this.
+# On first run the owner account + one-time password print to data/server.log.
 set -euo pipefail
 cd "$(dirname "$0")/.."
 lsof -ti:8777 | xargs kill 2>/dev/null || true
@@ -7,6 +11,6 @@ sleep 1
 nohup /opt/homebrew/bin/uv run uvicorn server.main:app --host 0.0.0.0 --port 8777 \
   > data/server.log 2>&1 &
 sleep 2
-curl -sf -o /dev/null http://localhost:8777/api/operators \
+curl -sf -o /dev/null http://localhost:8777/api/health \
   && echo "MODELBEAST up: http://$( /Applications/Tailscale.app/Contents/MacOS/Tailscale ip -4 2>/dev/null || echo localhost):8777 (log: data/server.log)" \
   || { echo "FAILED to start — tail data/server.log:"; tail -20 data/server.log; exit 1; }

server/auth.py

@@ -40,6 +40,21 @@ def verify_pw(pw: str, pw_hash: str) -> bool:
         return False
 
 
+# a real hash to compare against when the user doesn't exist, so login timing is
+# the same whether or not the username is valid (defeats username enumeration)
+_DUMMY_HASH = bcrypt.hashpw(b"modelbeast-dummy", bcrypt.gensalt()).decode()
+
+
+def check_login(con, username: str, password: str) -> dict | None:
+    """Constant-ish-time credential check. Returns the user row or None."""
+    u = get_user_by_name(con, username)
+    if u and verify_pw(password, u["pw_hash"]):
+        return u
+    if not u:
+        verify_pw(password, _DUMMY_HASH)  # burn equivalent time
+    return None
+
+
 # -- auth secret (stored hidden in settings; never exposed via the API) -------
 def ensure_auth_secret(con) -> str:
     row = con.execute("SELECT value FROM settings WHERE key='auth_secret'").fetchone()
@@ -56,14 +71,23 @@ def _serializer(con) -> URLSafeTimedSerializer:
 
 
 def make_session(con, user_id: str) -> str:
-    return _serializer(con).dumps({"uid": user_id})
+    u = get_user(con, user_id)
+    epoch = u.get("session_epoch", 0) if u else 0
+    return _serializer(con).dumps({"uid": user_id, "e": epoch})
 
 
 def read_session(con, token: str) -> str | None:
+    """Return the user id iff the signature is valid, unexpired, and the embedded
+    session epoch still matches (password change / logout-all bumps the epoch)."""
     try:
-        return _serializer(con).loads(token, max_age=SESSION_MAX_AGE).get("uid")
+        payload = _serializer(con).loads(token, max_age=SESSION_MAX_AGE)
     except (BadSignature, SignatureExpired, Exception):
         return None
+    uid = payload.get("uid")
+    u = get_user(con, uid) if uid else None
+    if not u or payload.get("e", 0) != u.get("session_epoch", 0):
+        return None
+    return uid
 
 
 # -- API tokens ----------------------------------------------------------------
@@ -121,7 +145,9 @@ def create_user(con, username: str, password: str, role: str = "guest",
 def set_password(con, user_id: str, password: str) -> None:
     if len(password) < MIN_PW_LEN:
         raise ValueError(f"password must be >= {MIN_PW_LEN} characters")
-    con.execute("UPDATE users SET pw_hash = ? WHERE id = ?", (hash_pw(password), user_id))
+    # bump session_epoch so every existing session/cookie for this user is revoked
+    con.execute("UPDATE users SET pw_hash = ?, session_epoch = session_epoch + 1 WHERE id = ?",
+                (hash_pw(password), user_id))
     con.commit()
 
 

server/db.py

@@ -61,6 +61,7 @@ MIGRATIONS = [
     ("jobs", "asset_ids", "TEXT NOT NULL DEFAULT '[]'"),
     ("jobs", "user_id", "TEXT"),      # nullable; legacy rows = owner-era
     ("assets", "user_id", "TEXT"),
+    ("users", "session_epoch", "INTEGER NOT NULL DEFAULT 0"),  # bump to revoke sessions
 ]
 
 

server/main.py

@@ -53,19 +53,27 @@ def _with_usernames(con, rows: list[dict]) -> list[dict]:
     return rows
 
 
+# -- health (public, unauthenticated — for serve.sh + proxy monitoring) --------
+@app.get("/api/health")
+def health():
+    return {"ok": True, "operators": len(runner.operators)}
+
+
 # -- auth / session ------------------------------------------------------------
 @app.post("/api/login")
 async def login(payload: dict, request: Request, response: Response):
-    ip = request.headers.get("x-forwarded-for", "").split(",")[0].strip() \
+    # Key the limiter on the DIRECT peer (Caddy) — not the client-spoofable
-        or (request.client.host if request.client else "?")
+    # X-Forwarded-For — plus a per-username bucket so the owner account can't be
-    if not auth.rate_ok(ip):
+    # brute-forced regardless of a rotated XFF header.
+    ip = request.client.host if request.client else "?"
+    username = (payload.get("username") or "").strip().lower()
+    if not auth.rate_ok(f"ip:{ip}") or not auth.rate_ok(f"user:{username}"):
         raise HTTPException(429, "too many attempts — wait a minute")
     con = _con(request)
-    username = (payload.get("username") or "").strip().lower()
     password = payload.get("password") or ""
-    u = auth.get_user_by_name(con, username)
+    u = auth.check_login(con, username, password)  # constant-time, no user enumeration
-    if not u or not auth.verify_pw(password, u["pw_hash"]):
+    if not u:
-        auth.note_fail(ip)
+        auth.note_fail(f"ip:{ip}"); auth.note_fail(f"user:{username}")
         raise HTTPException(401, "invalid username or password")
     token = auth.make_session(con, u["id"])
     secure = request.headers.get("x-forwarded-proto", "http") == "https"
@@ -94,10 +102,11 @@ def get_users(request: Request, owner=Depends(auth.require_owner)):
 @app.post("/api/users")
 def add_user(payload: dict, request: Request, owner=Depends(auth.require_owner)):
     try:
+        maxj = int(payload.get("max_active_jobs", 4))
         u = auth.create_user(_con(request), payload.get("username", ""),
                              payload.get("password", ""), role=payload.get("role", "guest"),
-                             max_active_jobs=int(payload.get("max_active_jobs", 4)))
+                             max_active_jobs=maxj)
-    except ValueError as e:
+    except (ValueError, TypeError) as e:
         raise HTTPException(400, str(e))
     return auth.public_user(u)
 
@@ -114,8 +123,11 @@ def patch_user(username: str, payload: dict, request: Request, owner=Depends(aut
         except ValueError as e:
             raise HTTPException(400, str(e))
     if "max_active_jobs" in payload:
-        con.execute("UPDATE users SET max_active_jobs = ? WHERE id = ?",
+        try:
-                    (int(payload["max_active_jobs"]), u["id"]))
+            maxj = int(payload["max_active_jobs"])
+        except (ValueError, TypeError):
+            raise HTTPException(400, "max_active_jobs must be an integer")
+        con.execute("UPDATE users SET max_active_jobs = ? WHERE id = ?", (maxj, u["id"]))
         con.commit()
     return auth.public_user(auth.get_user(con, u["id"]))
 
@@ -177,23 +189,43 @@ def put_settings(payload: dict, request: Request, owner=Depends(auth.require_own
 
 
 # -- assets --------------------------------------------------------------------
+def _can_see(user, obj) -> bool:
+    """owner sees everything; a guest sees only their own jobs/assets."""
+    return user["role"] == "owner" or obj.get("user_id") == user["id"]
+
+
+def _guest_may_run(user, op) -> bool:
+    """HARD RULE: guests are local-only. Blocked by TWO signals (defense in depth):
+    any declared requires_env (paid keys) AND the net/cloud lane. So a paid operator
+    that forgot to declare requires_env is still blocked."""
+    if user["role"] == "owner":
+        return True
+    return not op.get("requires_env") and op.get("resources") != "net"
+
+
 @app.get("/api/assets")
 def list_assets(user=Depends(auth.current_user), request: Request = None):
     con = _con(request)
-    return _with_usernames(con, store.list_assets(con))
+    rows = [a for a in store.list_assets(con) if _can_see(user, a)]
+    return _with_usernames(con, rows)
 
 
 @app.post("/api/assets")
 async def upload_asset(request: Request, file: UploadFile = File(...),
                        user=Depends(auth.current_user)):
-    cl = request.headers.get("content-length")
+    # stream in bounded chunks so a missing/spoofed Content-Length can't blow memory
-    if cl and int(cl) > MAX_UPLOAD:
+    chunks, total = [], 0
+    while True:
+        chunk = await file.read(1024 * 1024)
+        if not chunk:
+            break
+        total += len(chunk)
+        if total > MAX_UPLOAD:
             raise HTTPException(413, "file too large (1 GB max)")
-    data = await file.read()
+        chunks.append(chunk)
+    data = b"".join(chunks)
     if not data:
         raise HTTPException(400, "empty file")
-    if len(data) > MAX_UPLOAD:
-        raise HTTPException(413, "file too large (1 GB max)")
     return store.register_upload(_con(request), file.filename or "unnamed", data, user_id=user["id"])
 
 
@@ -203,7 +235,7 @@ def delete_asset(asset_id: str, user=Depends(auth.current_user), request: Reques
     asset = store.get_asset(con, asset_id)
     if not asset:
         raise HTTPException(404, "no such asset")
-    if user["role"] != "owner" and asset.get("user_id") != user["id"]:
+    if not _can_see(user, asset):
         raise HTTPException(403, "not your asset")
     store.delete_asset(con, asset_id)
     return {"ok": True}
@@ -215,10 +247,12 @@ def asset_file(asset_id: str, member: str | None = None,
     asset = store.get_asset(_con(request), asset_id)
     if not asset:
         raise HTTPException(404, "no such asset")
+    if not _can_see(user, asset):
+        raise HTTPException(403, "not your asset")
     path = Path(asset["path"])
     if member:
         target = (path / member).resolve()
-        if not str(target).startswith(str(path.resolve())) or not target.is_file():
+        if not target.is_relative_to(path.resolve()) or not target.is_file():
             raise HTTPException(404, "no such member")
         return FileResponse(target)
     if path.is_dir():
@@ -231,13 +265,15 @@ def asset_file(asset_id: str, member: str | None = None,
 @app.get("/api/jobs")
 def list_jobs(user=Depends(auth.current_user), request: Request = None):
     con = _con(request)
-    return _with_usernames(con, runner.list_jobs(con))
+    rows = [j for j in runner.list_jobs(con) if _can_see(user, j)]
+    return _with_usernames(con, rows)
 
 
 @app.get("/api/jobs/recent")
 def recent_jobs(user=Depends(auth.current_user), request: Request = None):
     con = _con(request)
-    return sysinfo.recent_jobs(con, runner)
+    uid = None if user["role"] == "owner" else user["id"]
+    return sysinfo.recent_jobs(con, runner, user_id=uid)
 
 
 @app.post("/api/jobs")
@@ -247,8 +283,7 @@ async def create_job(payload: dict, user=Depends(auth.current_user), request: Re
     if operator not in runner.operators:
         raise HTTPException(400, f"unknown operator: {operator}")
     op = runner.operators[operator]
-    # HARD RULE: guests run local models only — enforced here regardless of UI
+    if not _guest_may_run(user, op):
-    if op.get("requires_env") and user["role"] != "owner":
         raise HTTPException(403, "guests run local models only")
     if user["role"] != "owner" and runner.active_job_count(con, user["id"]) >= user["max_active_jobs"]:
         raise HTTPException(429, "too many active jobs — wait for some to finish")
@@ -256,8 +291,11 @@ async def create_job(payload: dict, user=Depends(auth.current_user), request: Re
     if not asset_ids:
         asset_ids = [payload["asset_id"]] if payload.get("asset_id") else []
     for aid in asset_ids:
-        if not store.get_asset(con, aid):
+        a = store.get_asset(con, aid)
+        if not a:
             raise HTTPException(400, f"unknown asset: {aid}")
+        if not _can_see(user, a):
+            raise HTTPException(403, "not your asset")
     params = payload.get("params") or {}
     job = runner.create_job(con, operator, asset_ids, params, user_id=user["id"])
     await runner.broadcast({"type": "job", "job": job})
@@ -295,8 +333,10 @@ async def retry_job(job_id: str, user=Depends(auth.current_user), request: Reque
     con = _con(request)
     job = _own_job_or_owner(con, job_id, user)
     op = runner.operators.get(job["operator"], {})
-    if op.get("requires_env") and user["role"] != "owner":
+    if not _guest_may_run(user, op):
         raise HTTPException(403, "guests run local models only")
+    if user["role"] != "owner" and runner.active_job_count(con, user["id"]) >= user["max_active_jobs"]:
+        raise HTTPException(429, "too many active jobs — wait for some to finish")
     new = runner.create_job(con, job["operator"], job.get("asset_ids") or [],
                             job["params"], user_id=user["id"])
     await runner.broadcast({"type": "job", "job": new})
@@ -341,7 +381,8 @@ async def watch_inbox():
             print(f"[inbox] {e}")
 
 
-# -- websocket (auth: session cookie or ?token=) -------------------------------
+# -- websocket (auth: session cookie only — never a token in the query string,
+#    which would leak into proxy/access logs) ----------------------------------
 @app.websocket("/ws")
 async def ws(websocket: WebSocket):
     con = websocket.app.state.con
@@ -351,20 +392,16 @@ async def ws(websocket: WebSocket):
         uid = auth.read_session(con, tok)
         if uid:
             user = auth.get_user(con, uid)
-    if not user:
-        qtok = websocket.query_params.get("token")
-        if qtok:
-            user = auth.user_by_token(con, qtok)
     if not user:
         await websocket.close(code=1008)  # policy violation
         return
     await websocket.accept()
-    runner.subscribers.add(websocket)
+    runner.subscribers[websocket] = user
     try:
         while True:
             await websocket.receive_text()
     except WebSocketDisconnect:
-        runner.subscribers.discard(websocket)
+        runner.subscribers.pop(websocket, None)
 
 
 # -- frontend (built) — public so the SPA can load and show the login screen ----

server/runner.py

@@ -27,7 +27,7 @@ LANE_LIMITS = {"gpu": 1, "cpu": 3, "net": 6}
 class Runner:
     def __init__(self):
         self.queue: asyncio.Queue[str] = asyncio.Queue()
-        self.subscribers: set = set()
+        self.subscribers: dict = {}    # websocket -> user dict (for per-user scoping)
         self.operators: dict[str, dict] = {}
         self.lanes: dict[str, asyncio.Semaphore] = {}
         self.procs: dict = {}          # job_id -> subprocess
@@ -37,14 +37,20 @@ class Runner:
 
     # -- pubsub -------------------------------------------------------------
     async def broadcast(self, message: dict):
+        # "job" events carry a payload injected straight into the client's list, so
+        # scope them: a guest only receives their own jobs (owner receives all).
+        # "assets_changed" just triggers a per-user re-fetch, so send to everyone.
+        job_uid = message.get("job", {}).get("user_id") if message.get("type") == "job" else None
         dead = []
-        for ws in self.subscribers:
+        for ws, u in list(self.subscribers.items()):
+            if job_uid is not None and u.get("role") != "owner" and job_uid != u.get("id"):
+                continue
             try:
                 await ws.send_json(message)
             except Exception:
                 dead.append(ws)
         for ws in dead:
-            self.subscribers.discard(ws)
+            self.subscribers.pop(ws, None)
 
     def lane_of(self, operator: str) -> str:
         op = self.operators.get(operator, {})
@@ -172,7 +178,9 @@ class Runner:
 
         cur_settings = self.get_settings()
         env = os.environ.copy()
-        env.update(settings_mod.env_from_settings(cur_settings))
+        # least privilege: only this operator's declared paid keys (+ HF/infra) reach
+        # the subprocess, so guest local jobs never get the owner's fal/OpenRouter keys
+        env.update(settings_mod.env_for_operator(cur_settings, op.get("requires_env")))
         secrets = settings_mod.secret_values(cur_settings)
 
         missing = [e for e in op.get("requires_env", []) if not env.get(e)]

server/settings.py

@@ -71,6 +71,26 @@ def env_from_settings(settings: dict) -> dict:
     return env
 
 
+def env_for_operator(settings: dict, requires_env) -> dict:
+    """Least-privilege env for an operator subprocess: non-secret infra vars + the
+    HF token (local ML ops need it for gated weights) ALWAYS; paid/secret API keys
+    ONLY if the operator declares them in requires_env. So a guest's local job never
+    receives the owner's fal/OpenRouter keys even in its environment."""
+    requires_env = set(requires_env or [])
+    env = {}
+    for key, env_name in ENV_MAP.items():
+        value = settings.get(key)
+        if not value:
+            continue
+        # paid secret keys are gated by requires_env; hf_token + non-secret infra always
+        if key in SECRET_KEYS and key != "hf_token" and env_name not in requires_env:
+            continue
+        env[env_name] = value
+    if settings.get("hf_token"):
+        env["HUGGING_FACE_HUB_TOKEN"] = settings["hf_token"]
+    return env
+
+
 def secret_values(settings: dict) -> list[str]:
     return [settings[k] for k in SECRET_KEYS if settings.get(k)]
 

server/store.py

@@ -23,10 +23,17 @@ def kind_of(name: str) -> str:
     return "other"
 
 
+def safe_name(name: str) -> str:
+    """Strip any directory components from a user-supplied filename (path-traversal
+    guard) so the stored file can only ever land inside its own asset dir."""
+    base = Path(name or "").name
+    return base if base and base not in (".", "..") else "unnamed"
+
+
 def register_file(con, src: Path, name: str | None = None, parent_job: str | None = None,
                   move: bool = False, meta: dict | None = None, user_id: str | None = None) -> dict:
     """Copy/move a file (or directory) into the asset store and record it."""
-    name = name or src.name
+    name = safe_name(name or src.name)
     asset_id = db.new_id()
     dest_dir = ASSETS_DIR / asset_id
     dest_dir.mkdir(parents=True, exist_ok=True)
@@ -59,6 +66,7 @@ def register_file(con, src: Path, name: str | None = None, parent_job: str | Non
 
 
 def register_upload(con, name: str, data: bytes, user_id: str | None = None) -> dict:
+    name = safe_name(name)
     asset_id = db.new_id()
     dest_dir = ASSETS_DIR / asset_id
     dest_dir.mkdir(parents=True, exist_ok=True)

server/sysinfo.py

@@ -106,11 +106,17 @@ def snapshot(con, runner) -> dict:
     }
 
 
-def recent_jobs(con, runner, limit: int = 20) -> list[dict]:
+def recent_jobs(con, runner, limit: int = 20, user_id: str | None = None) -> list[dict]:
     cutoff = time.time() - 86400
+    if user_id is None:  # owner: all users
         rows = con.execute(
             "SELECT * FROM jobs WHERE status IN ('done','error','cancelled') "
             "AND finished_at >= ? ORDER BY finished_at DESC LIMIT ?", (cutoff, limit)).fetchall()
+    else:  # guest: only their own
+        rows = con.execute(
+            "SELECT * FROM jobs WHERE status IN ('done','error','cancelled') "
+            "AND finished_at >= ? AND user_id = ? ORDER BY finished_at DESC LIMIT ?",
+            (cutoff, user_id, limit)).fetchall()
     umap = {u_id: name for u_id, name in
             con.execute("SELECT id, username FROM users").fetchall()}
     out = []

tests/smoke.sh

@@ -91,12 +91,18 @@ c=$(code -b "$GJAR" -X POST "$B/api/jobs" -H 'Content-Type: application/json' -d
 c=$(code -b "$GJAR" "$B/api/settings"); [ "$c" = 403 ] && ok "guest settings → 403" || bad "guest settings ($c)"
 c=$(code -b "$GJAR" -X POST "$B/api/users" -H 'Content-Type: application/json' -d '{"username":"x","password":"yyyyyy"}')
 [ "$c" = 403 ] && ok "guest cannot create users → 403" || bad "guest user-create ($c)"
-# guest CAN run a local operator
+# guest uploads their OWN asset and runs a local op on it → 200
+GVID=$(curl -s -b "$GJAR" -F "file=@$TMP/clip.mp4" "$B/api/assets" | j "d['id']")
+c=$(code -b "$GJAR" -X POST "$B/api/jobs" -H 'Content-Type: application/json' -d '{"operator":"ffprobe","asset_id":"'"$GVID"'"}')
+[ "$c" = 200 ] && ok "guest runs local op on OWN asset" || bad "guest own-asset run ($c)"
+# per-user isolation: guest CANNOT use / see / delete / download the owner's asset
 FFPAY='{"operator":"ffprobe","asset_id":"'"$VID"'"}'
 c=$(code -b "$GJAR" -X POST "$B/api/jobs" -H 'Content-Type: application/json' -d "$FFPAY")
-[ "$c" = 200 ] && ok "guest can run local ffprobe" || bad "guest local blocked ($c)"
+[ "$c" = 403 ] && ok "guest cannot use owner asset → 403" || bad "guest used owner asset ($c)"
-# guest cannot delete owner's asset
 c=$(code -b "$GJAR" -X DELETE "$B/api/assets/$GLB"); [ "$c" = 403 ] && ok "guest cannot delete owner asset → 403" || bad "guest deleted owner asset ($c)"
+c=$(code -b "$GJAR" "$B/api/assets/$GLB/file"); [ "$c" = 403 ] && ok "guest cannot download owner asset → 403" || bad "guest downloaded owner asset ($c)"
+GSEE=$(curl -s -b "$GJAR" "$B/api/assets" | j "any(a['id']=='$GLB' for a in d)")
+[ "$GSEE" = False ] && ok "guest asset list excludes owner's" || bad "guest sees owner assets"
 # per-user cap: capped user with max_active_jobs=0 → any job 429
 curl -s -b "$OJAR" -X POST "$B/api/users" -H 'Content-Type: application/json' -d '{"username":"capped","password":"cappedpw"}' >/dev/null
 curl -s -b "$OJAR" -X PATCH "$B/api/users/capped" -H 'Content-Type: application/json' -d '{"max_active_jobs":0}' >/dev/null
@@ -104,10 +110,22 @@ curl -s -c "$CJAR" -X POST "$B/api/login" -H 'Content-Type: application/json' -d
 c=$(code -b "$CJAR" -X POST "$B/api/jobs" -H 'Content-Type: application/json' -d "$FFPAY")
 [ "$c" = 429 ] && ok "per-user cap → 429" || bad "cap not enforced ($c)"
 
+say "hardening (from security review)"
+# upload filename path-traversal is sanitized to a basename
+TRAV=$(curl -s -b "$OJAR" -F "file=@$TMP/clip.mp4;filename=../../../evil.mp4" "$B/api/assets" | j "d['name']")
+[ "$TRAV" = "evil.mp4" ] && ok "upload filename sanitized to basename" || bad "filename traversal ($TRAV)"
+# password change revokes the existing session (session_epoch bump)
+curl -s -b "$OJAR" -X PATCH "$B/api/users/testmate" -H 'Content-Type: application/json' -d '{"password":"newpass123"}' >/dev/null
+c=$(code -b "$GJAR" "$B/api/me"); [ "$c" = 401 ] && ok "password change revokes old session" || bad "stale session still valid ($c)"
+# health endpoint is public (for serve.sh / proxy)
+c=$(code "$B/api/health"); [ "$c" = 200 ] && ok "/api/health public" || bad "health ($c)"
+
 say "dashboard"
 SHAPE=$(curl -s -b "$OJAR" "$B/api/system" | j "'gpu' in d and 'lanes' in d")
 [ "$SHAPE" = True ] && ok "/api/system shape" || bad "system shape"
-c=$(code -b "$GJAR" "$B/api/system"); [ "$c" = 200 ] && ok "guest can view dashboard" || bad "guest dashboard ($c)"
+CJAR2="$TMP/mate2.cj"
+curl -s -c "$CJAR2" -X POST "$B/api/login" -H 'Content-Type: application/json' -d '{"username":"testmate","password":"newpass123"}' >/dev/null
+c=$(code -b "$CJAR2" "$B/api/system"); [ "$c" = 200 ] && ok "guest can view dashboard" || bad "guest dashboard ($c)"
 
 printf "\n\033[1mResult: %d passed, %d failed\033[0m\n" "$pass" "$fail"
 [ "$fail" -eq 0 ]