Desktop render, upload, and token-boundary batch¶
Signal: The 2026-05-08 19:15 UTC advisory scan added a mixed but durable set of boundary failures across Electron renderers, model/admin UIs, backup portals, module updaters, parser resource limits, and enrollment handshakes.
Advisory cluster¶
- SiYuan Electron renderer RCE via tooltip XSS — GHSA-25rp-h46x-2hjm: incomplete fix for CVE-2026-34585;
decodeURIComponent-drivenaria-labeltooltip rendering can reach Electron RCE. No patched version listed at advisory time. - open-webui stored XSS via model description — GHSA-gf5m-wcrh-7928: npm/pip
open-webui <= 0.8.12; patch to0.9.0+. - Rdiffweb access-token impersonation — GHSA-v4gp-hf5j-4566:
rdiffweb < 2.10.6allowed any valid or stolen token to act as other users. Patch to2.10.6+. - OpenSTAManager arbitrary module upload — GHSA-rm34-fg4m-39mw:
devcode-it/openstamanager <= 2.10-beta; no patched version listed at advisory time. - HashiCorp Boundary worker enrollment TLS handshake DoS — GHSA-7x9r-wcgg-w86f: patch Boundary to
0.19.5+,0.20.3+, or0.21.3+depending on line. - justhtml parser DoS hardening — GHSA-r8cj-3554-33mr:
justhtml < 1.18.0; patch to1.18.0+. - phpseclib ASN.1 OID length duplicate/canonicalization update — GHSA-f2qx-66wf-wvvx, withdrawn duplicate GHSA-jr22-8qgm-4q87: patch
phpseclib/phpseclibto1.0.23+,2.0.47+, or3.0.36+for this OID-length hardening path; see the newer parser/crypto phpseclib guidance for later patch floors where applicable.
Triage¶
- Prioritize Electron/desktop apps, AI/model admin UIs, backup portals, and module/plugin uploaders on developer or admin workstations.
- Patch open-webui and Rdiffweb immediately; isolate or disable OpenSTAManager module-update upload paths until a fixed release is available.
- For SiYuan, treat untrusted notebook content, imported docs, and synced workspaces as code-adjacent if Electron integration is enabled.
- For Boundary, rate-limit worker enrollment endpoints and look for handshake spikes or enrollment failures from repeated sources.
- For parser/resource issues, enforce input size, nesting, and time budgets before expensive decode/render steps.
Hunt prompts¶
- Stored model descriptions, notebook labels, tooltips, backup labels, module metadata, or admin UI fields containing encoded HTML, event handlers, SVG,
javascript:, ordecodeURIComponentgadget payloads. - Backup/admin actions where one user token performs operations or reads data for another account.
- Module update ZIPs, PHP files, or plugin archives uploaded outside expected maintenance windows.
- Boundary worker enrollment logs with repeated TLS failures, high CPU, or queue exhaustion.
- ASN.1/OID or HTML parser inputs that produce unusually high CPU, memory, or timeout rates.
Durable controls¶
- Keep renderer content sanitized at every sink, not only at initial import; tooltips, titles, ARIA labels, markdown previews, and model descriptions are executable UI surfaces.
- Scope access tokens to an immutable subject and re-check subject/resource authorization server-side on every request.
- Treat plugin/module upload as code deployment: require admin MFA, signed artifacts, extension allowlists, unpack path controls, and post-upload scanning.
- Budget parser work before allocation and make malformed input cheap to reject.
- Separate worker enrollment/control-plane handshakes from public or attacker-reachable network paths where possible.