Siyuan EUVD-2026-74836

Exploitation requires that the Siyuan instance be deployed in publish mode, which enables unauthenticated external access to the application's API. This is not the standard single-user local deployment model; publish mode must be explicitly enabled. The attacker needs only network access to the Siyuan web interface - no credentials, no prior knowledge of document structure, and no special client software. Limiting factors: instances running in default single-user local mode are not exposed; the disclosed information is limited to aggregate counts (not content); and the attacker must iteratively probe with specific search terms, so discovery is bounded by the quality of the attacker's guesses. There is no evidence this requires any non-default feature beyond publish mode itself.

The CVSS 4.0 vector (AV:N/AC:L/AT:N/PR:N/UI:N/VC:L) scores 6.9 and accurately characterizes this as a low-confidentiality, network-accessible, unauthenticated issue - the score reflects the limited nature of what leaks (metadata counts rather than raw content). Real-world risk is moderate and deployment-context-dependent: instances operated entirely in private mode (no publish-mode exposure) are not affected, substantially narrowing the population at risk. For operators running Siyuan in publish mode with sensitive unpublished content - from personal researchers to small teams - the oracle allows systematic enumeration of whether specific keywords, names, or phrases appear in hidden notes, which can be meaningful even without content access. No KEV listing and no public exploit code are identified, so mass exploitation is unlikely; the EPSS signal is unavailable. The absence of integrity or availability impact is consistent with the description. Patching to v3.8.2 is straightforward, so even a mid-sized internal team that has deployed this in a shared publish-mode configuration should prioritize the upgrade without treating this as a crisis-level event.

An attacker discovers a Siyuan instance running in publish mode and submits a series of targeted POST requests to /api/search/fullTextSearchBlock with carefully chosen search terms - employee names, project codenames, or financial keywords. The responses omit actual block content from private documents but return match counts; by iterating over many terms, the attacker builds a keyword map of what exists in hidden notes. No exploit code is required - standard HTTP tooling suffices given the unauthenticated network-accessible endpoint.