Vulnerability  ·  2026-11-20

DOMShell MCP: domshell_eval is misclassified as read-tier, bypassing the write authorization boundary

VulnerabilityLow impactGlobalCVE-2026-75518
The commit reclassified eval/js into the write tier and updates tool descriptions and MCP instructions, noting that no new capability was exposed (same write access already existed via domshell_js with --allow-write) but the tier misclassification misrepresented the granted capability in MCP hosts, tier flags and audit logs.
In an agentic browser-automation tool, a read/write-tier misclassification means the authorization surface shown to the model (and recorded in audit) understates the capability actually being granted, which can let an agent mutate browser/application state under the guise of read-only access.
domshell_eval is classified as a read-tier operation that runs without --allow-write, but it shares the same CDP Runtime.evaluate handler (no throwOnSideEffect gate) as the write-tier domshell_js, so any JS expression can mutate DOM/window/global state without the write flag being granted — an authorization-boundary bypass (CWE-863 class) that MCP hosts and audit logs misrepresent.
apireno/DOMShell MCP server below 2.0.8 (domshell_eval / eval / js)
Upgrade to DOMShell MCP server 2.0.8 or later, which moves eval into WRITE_COMMANDS and changes the domshell_eval annotation from read to write so hosts render the correct approval UI.
DOMShell MCP eval-tier fix commitNVD — CVE-2026-75518
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →