Skip to content

~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.

3.9 KB
Raw Blame

🔒 Security

  • Notification URL validation now rejects Apprise query options that can select an unvalidated secondary destination or resource (template, redirect, and mail-specific SMTP/PGP/WKD options) before any hostname parsing or DNS lookup. A template option is an unrestricted Apprise attachment, so discord://id/token?template=/etc/passwd previously made the server read a local file and ship it to the user's webhook. Query-key parsing mirrors Apprise's own parse_qsd rules (& and ; separators, +-as-space, percent-decoding, duplicate keys) so encoded or semicolon-separated names cannot bypass the policy, and a malformed percent escape in a key is rejected outright. The published dependency now declares apprise>=1.12.0,<2 — 1.12.0 is the reviewed parser floor — and a test diffs LDR's canonical keys against the installed Apprise parser so the contract cannot drift silently. Notification security documentation and compatibility tests now cover mode-dependent destinations and the current direct scheme-prefix gaps. (#5085)

  • The notification test endpoint (/api/notifications/test-url) now validates every URL in a comma/space-separated multi-URL string individually. Previously the input was validated as a single URL — commas are legal URL characters, so only the leading entry was checked while Apprise's own scheme-aware splitter still registered and notified the rest, letting an unvetted destination (e.g. a cloud-metadata address) receive the test notification. The number of parsed targets is now also capped, and the test is refused outright if Apprise registers more services than were validated. (#5120)

  • Closed an SSRF hole in the news subscription API. A custom LLM endpoint (custom_endpoint) supplied when creating or updating a subscription is fetched server-side, and the /news/api blueprint — the one the app actually calls — accepted it without validation. The guard existed only on the /api/news blueprint, which nothing routes to; it is now applied where requests actually arrive, on both POST /news/api/subscribe and PUT /news/api/subscriptions/<id>, before anything is written or any research thread is spawned. A wrong-typed custom_endpoint (a number, a list) is now a 400 rather than a 500 or a silently stored value.

    The guard is a denylist, not an allowlist: it blocks cloud metadata, link-local addresses and non-HTTP schemes, but loopback and private-network addresses stay allowed on purpose, because that is how people point this at Ollama, LM Studio or vLLM. Other internal targets — including the app's own admin port and a Docker daemon on the bridge address — therefore remain reachable through this field by design. The check also validates the submitted URL only: it does not pin the resolved IP, and a redirect from an allowed host to a metadata address would still be followed by the LLM client, which is the one case the metadata denylist is meant to stop.

    GET /news/api/feed additionally rejects a subscription_id that is neither a UUID nor the documented all sentinel, and GET /news/api/subscriptions/<id>/history rejects a path id that is not a UUID. Those are the two ids that reach a LIKE pattern rather than an equality filter. Both checks are defence in depth rather than fixes: the LIKE queries already escape % and _ with an explicit escape character, so neither was exploitable. Subscription ids are generated as UUID4s, and the sibling /api/news blueprint has always applied the same check to its history route. The remaining subscription routes match their path id by equality, so they need no such check.

  • Reject ambiguous legacy numeric IPv4 host forms before DNS resolution in webhook and notification URL validation.