# SiYuan: the agent chat panel renders the last reply as raw HTML

> SiYuan sanitizes the streaming preview of an agent reply but not the final render, turning a model's answer into code execution.

- **Severity:** Medium: Real and reproducible, but needs unusual setup, user help or a narrow audience. Plan a fix.
- **Category:** Vulnerability
- **Published by:** SiYuan
- **Disclosed:** 2026-11-11
- **CVSS:** 7.5
- **Affects:** SiYuan
- **Research by:** president-xd
- **Primary source:** https://github.com/siyuan-note/siyuan/security/advisories/GHSA-3rvg-qxwg-f3c6
- **Page:** https://theinjection.dev/items/siyuan-agent-chat-panel-html-rce/

SiYuan's AI agent chat panel renders the final assistant message as unsanitized HTML, so an injected event handler can run operating-system commands in the desktop app's local-kernel mode. The maintainer scores it 7.5.

## What happened

SiYuan published advisory GHSA-3rvg-qxwg-f3c6 on November 11, 2026 for a DOM cross-site scripting flaw in its AI agent chat panel. The reporter is credited as president-xd. Versions 3.7.0 through 3.8.10 of the siyuan npm package are affected, and the advisory names v3.8.11 as the patched version.

## How it works

The chat panel builds its Markdown engine without HTML sanitization enabled, then assigns the rendered assistant output to innerHTML. A second engine instance sanitizes the streaming preview, so a user sees clean text while the reply streams and the unsanitized body replaces it when the stream ends, stops or errors. Block-level raw HTML in a reply keeps its inline event-handler attributes, and attributes supplied through kramdown attribute-list syntax survive even where sanitization is on. In desktop local-kernel mode the window runs with Node.js integration enabled and no active Content-Security-Policy, so a handler can call Node APIs.

## Who is affected

SiYuan desktop users in local-kernel mode carry the code-execution risk, per the advisory, while remote-kernel mode applies a Content-Security-Policy without unsafe-inline. Two routes reach the sink: a model reproducing attacker-influenced markup it read, which the advisory calls probabilistic, or a configured model endpoint returning the markup directly. Saved sessions re-render through the same path, so a payload fires again whenever the session is reopened.

## What to do

Treat SiYuan's agent panel as untrusted until a stable release carries the fix, because only v3.8.11 alpha builds are published and v3.8.10 remains the latest stable release. Delete saved agent sessions that could hold a payload, since reopening one runs it again, and point the panel only at model endpoints you control.

## Quick facts

- **Affected:** siyuan (npm) 3.7.0 through 3.8.10
- **Fixed in:** Advisory names v3.8.11; only v3.8.11 alpha pre-releases are published, with v3.8.10 the latest stable release
- **CVSS:** 7.5, scored by the maintainer in the advisory
- **CVE:** None assigned
- **Found by:** president-xd
- **Attack type:** DOM cross-site scripting reached through the model's reply
- **Exposed configuration:** Desktop local-kernel mode, where Node.js integration is on and no Content-Security-Policy applies

## Questions

### Is there a stable SiYuan release that fixes the agent chat panel flaw?

SiYuan's advisory names v3.8.11 as the patched version, but the project's releases page shows only v3.8.11 alpha pre-releases, the most recent published on November 16, 2026. Version 3.8.10, released November 10, is still marked as the latest stable release and sits inside the affected range. Anyone waiting for a stable fix has nothing to install yet.

### Which SiYuan configuration is exposed to code execution?

Desktop local-kernel mode is the exposed configuration, according to the advisory, because that window runs with Node.js integration enabled, context isolation disabled, web security disabled and no active Content-Security-Policy. An injected event handler can therefore call Node APIs and run operating-system commands with the user's privileges. Remote-kernel mode applies a policy without unsafe-inline.

### How would an attacker get markup into a SiYuan assistant reply?

The advisory describes two routes into the agent's reply. Indirect prompt injection is the first: the agent reads attacker-influenced material and the model reproduces the malicious markup, a route the advisory calls probabilistic. A model endpoint the user configured, returning the markup directly, is the second and deterministic route. Saved sessions then re-execute the payload whenever they are reopened.

## Sources

1. [SiYuan Advisory](https://github.com/siyuan-note/siyuan/security/advisories/GHSA-3rvg-qxwg-f3c6)
2. [SiYuan Releases](https://github.com/siyuan-note/siyuan/releases)
