CVE-2026-66421

OpenClaw Agent Dashboard (vendor tugcantopaloglu) contains a stored cross-site scripting vulnerability (CWE-79) in how it renders agent transcript messages. An unauthenticated remote attacker can submit a crafted message through the sessions API — for example an img tag with an onerror attribute that fits within the 60-character rendering budget — and the payload is stored in the session transcript and later interpolated unsanitized into innerHTML on the dashboard's default landing page. When an administrator opens that page, the injected JavaScript executes in the admin's browser session, allowing theft of session tokens and unauthorized calls to authenticated administrative endpoints, including modification of agent instruction files. Users running the affected dashboard are exposed wherever the landing page displays transcript content from sessions that untrusted parties can write messages into. The flaw is rated 8.8 (High) under CVSS 4.0, is not yet in CISA KEV, and has low predicted exploitation probability (EPSS 0.5%, percentile 43), but two public proof-of-concept references exist (a GitHub PoC and a VulnCheck advisory), so it should be treated as realistically exploitable even though in-the-wild exploitation is not confirmed.

What to do: Upgrade to the latest OpenClaw Dashboard release as soon as a patched version is published (no fixed version number is stated in the disclosure); as interim mitigations, restrict which parties can post messages into agent sessions, limit dashboard access to trusted networks, and enforce a Content-Security-Policy that blocks inline scripts and event handlers. Administrators should also check session transcripts for HTML-bearing messages (e.g., img tags with onerror handlers), revoke and rotate exposed dashboard/admin session tokens, and review agent instruction files for unauthorized modifications.