The CVSS 8.6 rating for CVE-2026-72804 captures content disclosure from SiYuan's graph endpoints, but it understates the actual risk model. The getGraph and getLocalGraph endpoints expose the complete reference topology of password-protected documents to unauthenticated users — that's a relational map, not just block content, and the distinction matters.
In SiYuan, documents have a dual surface: the content itself and the graph showing how blocks reference each other. An attacker who maps which protected blocks cite which protected blocks gains structural intelligence about argumentation, citation logic, and key nodes — information with independent value even without reading the text. This isn't theoretical. The graph is a map, and maps have leverage. A complete topology reveals central nodes (highly-cited blocks = core arguments), structural clusters, and critically — citation bridges linking protected content to unprotected content. That last detail is where the cascade detonates: the attacker can identify paths into protected material through unprotected intermediaries, then target those entry points for follow-on attacks like credential guessing or social engineering.
The 'publish-password' tier is the sharpest risk indicator. This tier exists specifically for documents intended for external sharing with select readers. The vulnerability voids that entire security contract — every external reader who received a publish-password link was implicitly assured the graph was safe to discuss, regardless of whether exploitation occurred.
Compounding the issue: the graph topology likely entered auxiliary stores long before any patch arrives. Search indexes, navigation caches, session preload data — these are forgotten layers where the topology survives. The patch closes the front door while the intelligence asset may already be cached elsewhere. For deployments with external-facing publish-password documents, the EPSS of 0.00255 is misleading — it models opportunistic mass exploitation, not targeted intelligence gathering. Your question isn't whether bots will scan for this; it's whether your SIEM catches enumeration patterns from an attacker building the complete map one document at a time.