github.com/siyuan-note/siyuan/kernel
Affected
- introduced
0, last affected0.0.0-20260421031503-96dfe0bea474(semver)
Fixed versions
None listed
Fix
No fixed version is listed.
Advisory
Also known as: CVE-2026-44588, GO-2026-4993
SiYuan: Electron Renderer RCE via decodeURIComponent-driven tooltip XSS in aria-label sink (incomplete fix for CVE-2026-34585)
Summary
The tooltip mouseover handler in app/src/block/popover.ts reads aria-label via getAttribute and passes it through decodeURIComponent before assigning to messageElement.innerHTML in app/src/dialog/tooltip.ts:41. The encoder used at the producer side, escapeAriaLabel in app/src/util/escape.ts:19-25, only handles HTML special characters (", ', <, literal <) — it leaves %XX URL-escapes untouched. So a doc title containing %3Cimg src=x onerror=...%3E round-trips through escapeAriaLabel and the HTML attribute layer unmodified. Then decodeURIComponent on the consumer side converts %3C to a literal < character (a real <, NOT a character reference). When that string is assigned to innerHTML, the HTML5 tokenizer enters TagOpenState on the literal <, parses the <img> element, and the onerror handler fires.
Because the renderer runs with nodeIntegration: true, contextIsolation: false, webSecurity: false (app/electron/main.js:407-411), require('child_process') is reachable from the injected handler, escalating to arbitrary code execution.
Doc titles, AV column names + descriptions, AV select options, file-tree tooltips all reach this sink because they're rendered into class="ariaLabel" elements with aria-label="${escapeAriaLabel(...)}". Doc title is the easiest plant — any user with create/rename access lands the payload, and the file survives .sy.zip round-trip without modification.
Why a "double HTML-decode" framing is wrong
A naïve reading of the chain might suggest that &lt; (the encoder output) decodes once at attribute-parse time to <, then a second time at innerHTML time to < — yielding a tag. That's incorrect and confirmed false by direct browser testing. Per the HTML5 spec, character references in DataState produce CHARACTER tokens (text), not TagOpenState transitions: the < resulting from a < reference is text data, never a tag-open delimiter. So the HTML-entity-only payload renders as visible literal text, not as a tag.
The actual bypass relies on decodeURIComponent producing a literal < (not a character reference) before innerHTML parses it. Literal < characters in the input stream DO trigger TagOpenState. URL encoding is the right vehicle because the encoder ignores %XX while the consumer chain decodes it.
Details
Encoder. app/src/util/escape.ts:19-25:
export const escapeAriaLabel = (html: string) => {
if (!html) { return html; }
return html.replace(/"/g, """).replace(/'/g, "'")
.replace(/</g, "&lt;").replace(/</g, "&lt;");
};
The four replacements only cover HTML special chars. %XX URL escapes are not touched.
Source — search-result rendering. app/src/search/util.ts:1406:
<span class="b3-list-item__text ariaLabel" ... aria-label="${escapeAriaLabel(title)}">${escapeGreat(title)}</span>
Same pattern at :1448, protyle/render/av/blockAttr.ts:205, protyle/render/av/col.ts:134, protyle/render/av/select.ts:36, search/unRef.ts:113. The title is built from getNotebookName(item.box) + getDisplayName(item.hPath, false) (line 1398). The hPath returned by /api/search/fullTextSearchBlock carries the user-set doc title verbatim — %XX URL-escapes pass through, only HTML special chars are entity-encoded by the kernel.
Consumer. app/src/block/popover.ts:33,144:
let tip = aElement.getAttribute("aria-label") || ""; // literal stored attribute value
// ... branch logic that doesn't apply to plain search results ...
showTooltip(decodeURIComponent(tip), aElement, ...); // ← decodes %XX into raw chars
decodeURIComponent is presumably present to handle URL-encoded asset paths in some hyperlink tooltips, but it's applied unconditionally to every aria-label-sourced tip — that's what enables this bypass.
Sink. app/src/dialog/tooltip.ts:41:
messageElement.innerHTML = message; // ← HTML parser sees the now-decoded raw `<` and starts parsing tags
Decode-chain trace for in-memory title %3Cimg src=x onerror="alert('SiYuan')"%3E (URL-encoded < > ', literal "):
| step | result |
|------|--------|
| in-memory title | %3Cimg src=x onerror="alert('SiYuan')"%3E |
| escapeAriaLabel writes (only " and ' get encoded — neither appears here as raw chars when ' is %27) | %3Cimg src=x onerror="alert(%27SiYuan%27)"%3E |
| HTML attribute set: aria-label="..." ; browser one-decodes named entities when storing | in-DOM value = %3Cimg src=x onerror="alert(%27SiYuan%27)"%3E |
| getAttribute("aria-label") | %3Cimg src=x onerror="alert(%27SiYuan%27)"%3E (verbatim) |
| decodeURIComponent(tip) | `<img src=x onerror="alert('SiYuan')">` (real < ' > chars) |
| messageElement.innerHTML = … | HTML parser tokenizes raw <img>, creates element, fails to load src=x, fires onerror → JS runs |
Renderer + reachability. Renderer posture and auto-admin gates same as the AV-name advisory (Advisory 1): nodeIntegration:true, contextIsolation:false, webSecurity:false at app/electron/main.js:407-411; empty-AccessAuthCode local auto-admin at kernel/model/session.go:261-287; chrome-extension:// Origin allowlist at session.go:277.
Suggested fix
messageElement.innerHTML = message;
with
messageElement.textContent = message;
For tooltips that legitimately need markup (memo rendering, hyperlink preview cards), introduce an explicit {html: true} flag on showTooltip(...) and route the message through DOMPurify.sanitize(message) before assigning to innerHTML.
try/catch with a clear scope. Aria-label content is not URL-encoded by design; decoding it is a footgun that converts otherwise-safe attributes into pre-parsed HTML.app/src/util/escape.ts (escapeHtml, escapeAttr, escapeAriaLabel, escapeGreat) into one Lute.EscapeHTMLStr-equivalent that escapes &, <, >, ", '. Context-specific encoders without compile-time enforcement keep producing bug-class variants.contextIsolation: true with a preload bridge — caps every future renderer XSS at "DOM only," not RCE.---
Reproduction (copy-paste-ready)
Tested on Windows with SiYuan v3.6.5 (kernel + Electron) and Microsoft Edge as the offline parser-validation engine. Linux/macOS users substitute py with python3 and use any modern Chromium-based browser (Edge/Chrome/Brave) for the standalone validation step.
Prereqs
curl -s http://127.0.0.1:6806/api/system/version
# → {"code":0,"msg":"","data":"3.6.5"}
lsNotebooks returns a usable id. Pin variables: API=http://127.0.0.1:6806
NOTEBOOK_ID=$(curl -s -X POST $API/api/notebook/lsNotebooks \
-H 'Content-Type: application/json' -d '{}' \
| python -c 'import sys,json; print(json.load(sys.stdin)["data"]["notebooks"][0]["id"])')
echo "Using notebook: $NOTEBOOK_ID"Critical Basis: The source's own band: CRITICAL
0, last affected 0.0.0-20260421031503-96dfe0bea474 (semver)None listed
No fixed version is listed.
Fix commands are the export's own generic commands. Ojuju never runs them. Check them against your setup first.
Check your own computer: guard packages check
Source: GitHub Advisory Database (reviewed advisories) · CC BY 4.0