unsloth/studio/backend/utils/preview_token.py
Daniel Han 80d3434d61
Studio: require signed capability tokens for /p preview links (#6666)
* Studio: require signed capability tokens for /p preview links

The public /p preview routes added in #6486 run model load and chat
generation as the admin user with no authentication. The only gate is the
preview ref, a deterministic outputs-root path (run or run/checkpoint) that
is guessable rather than secret. On a network-reachable Studio (--secure
tunnel or -H 0.0.0.0), an unauthenticated caller who guesses a ref can
consume GPU and probe a private fine-tuned checkpoint.

Make the share link an unguessable, revocable capability:

- Sign the canonical ref with a dedicated server-side secret (HMAC-SHA256,
  stored in app_secrets, independent of the JWT/login secret).
- Require a valid token on every /p chat, models, and page request before
  resolving a checkpoint or loading a model; missing or invalid tokens get a
  generic 404 so the surface never confirms a ref exists.
- Accept the token via ?k= (browser link and preview page) or
  Authorization: Bearer (OpenAI-compatible clients).
- Rotate the secret to revoke every outstanding link
  (POST /api/settings/preview-links/rotate).
- Clamp preview generation (max_tokens/max_completion_tokens <= 1024, n = 1)
  and set Referrer-Policy: no-referrer on the page so the token is not
  leaked via Referer.

Training history hands the authenticated owner the signed token, and the
copy-link button builds /p/{ref}?k={sig}.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: honor a lower caller token limit in the preview clamp

Codex review: when only the legacy max_tokens was sent, the clamp left
max_completion_tokens at the 1024 default, and _effective_max_tokens prefers
max_completion_tokens, so a request like max_tokens=16 could still generate up
to 1024 tokens. Derive one effective limit (max_completion_tokens wins, else the
legacy max_tokens) and pin both fields to it so a caller's lower limit is kept.

* Studio: add preview kill switch, rate limit, and revoke-links UI

Follow-ups to the /p preview capability work:

- Public-sharing kill switch: a persisted setting (default on) gates the public
  /p surface. When off, every preview request 404s even with a valid token, and
  the owner UI stops offering share links. GET/PUT /api/settings/preview-sharing;
  enforced in _verify_or_404.
- Per-IP rate limit on the preview chat route: a coarse in-process sliding-window
  limiter (20 req/min/IP) returns 429 + Retry-After before the GPU lock is taken.
  Client IP honors X-Forwarded-For only when UNSLOTH_STUDIO_TRUST_FORWARDED is
  set, matching the login limiter's trust model.
- Settings UI: a "Preview sharing" section with the public-sharing toggle and a
  "Revoke all preview links" button (confirm dialog) that rotates the secret.

Tests cover the kill switch (404 when off), the 429 path, the sliding window,
client-IP trust behavior, and the setting default.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: fix preview-fields sharing arg and refresh sigs after revoke

Codex review:
- P1: get_training_run_detail and update_training_run called _preview_fields
  with only output_dir after it gained a required sharing_on parameter, raising
  a 500 TypeError once get_run succeeded. Pass get_preview_sharing_enabled() at
  both sites; add a detail-endpoint regression test.
- P2: after rotating the preview secret from settings, the history grid still
  held stale preview_sig values, so a freshly copied link would 404. Emit
  emitTrainingRunsChanged() after a successful revoke so the grid refetches
  freshly signed refs.

* Studio: harden preview sharing controls (Codex review)

- Fail closed: a read failure on the preview-sharing kill switch now returns
  False instead of defaulting to enabled, so an unavailable settings DB can't
  reopen the public surface. A missing key still defaults to enabled.
- Per-IP rate limit behind the managed Cloudflare tunnel: client_ip now honors
  CF-Connecting-IP when the socket peer is loopback, so tunneled visitors are
  keyed by their real IP instead of collapsing onto the local cloudflared peer.
- GET /p no longer mints key/share_url when sharing is disabled; it returns
  sharing_enabled=false so clients don't distribute links that 404.
- Settings UI: toggling public sharing emits the training-runs-changed event so
  the history grid shows/hides Copy preview link without a manual refresh.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: harden preview rate limiter and IP keying (Opus review)

From a two-agent review of the PR:

- Rate limiter no longer evicts an active bucket when the table is full: a flood
  of distinct keys could otherwise cycle out a throttled bucket and reset its
  counter. Evict only aged-out buckets; if the table is full of live clients,
  fail closed (deny the new key) instead.
- client_ip keys on the rightmost (proxy-appended) X-Forwarded-For hop when the
  trust env is set; the leftmost is client-spoofable. Documented the
  append/overwrite-proxy assumption.
- _verify_or_404 checks the capability token before the kill-switch DB read, so
  unauthenticated /p spam can't be used as an unbounded settings-DB sink and the
  response is identical regardless of the sharing on/off state.

Tests: nested run/checkpoint happy path + wrong-ref rejection, the eviction
fail-closed behavior, and route-level coverage for the rotate / preview-sharing
settings endpoints.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
2026-06-25 21:40:48 -07:00

52 lines
2 KiB
Python

# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""HMAC capability tokens for public ``/p`` preview share links.
The preview ref (``run`` or ``run/checkpoint``) is a deterministic, guessable
outputs-root path, so it can't gate access on its own. We sign the canonical ref
with a dedicated server-side secret and require the resulting token on every
public preview request: guessing a ref no longer grants access, and rotating the
secret (``auth.storage.rotate_preview_link_secret``) revokes every link at once.
"""
from __future__ import annotations
import base64
import hashlib
import hmac
from typing import Optional
from auth.storage import get_or_create_preview_link_secret
# Versioned so the token format can evolve without silently honoring old shapes.
_PREVIEW_TOKEN_VERSION = "v1"
def _canonical_payload(ref: str) -> bytes:
# Sign the canonical ref only (never host/path) so links stay portable across
# localhost / LAN IP / tunnel host changes.
return f"preview:{_PREVIEW_TOKEN_VERSION}:{ref}".encode("utf-8")
def sign_preview_ref(ref: str) -> str:
"""Return the URL-safe HMAC capability token for a canonical preview ref."""
mac = hmac.new(
get_or_create_preview_link_secret(),
_canonical_payload(ref),
hashlib.sha256,
).digest()
return base64.urlsafe_b64encode(mac).rstrip(b"=").decode("ascii")
def verify_preview_ref(ref: str, token: Optional[str]) -> bool:
"""Constant-time check that ``token`` is a valid capability for ``ref``."""
if not token:
return False
# Compare as bytes: a non-ASCII token (e.g. a %-encoded query value) would make
# hmac.compare_digest on two str raise TypeError -> treat it as simply invalid.
try:
provided = token.encode("ascii")
except UnicodeEncodeError:
return False
return hmac.compare_digest(sign_preview_ref(ref).encode("ascii"), provided)