Friday, 20 November 2026
2 agent hacks today▼ 9 vs yesterday (11)

DOMShell tool let agents write to pages under a read-only label

A GitHub advisory published on 19 November 2026 says Apireno's domshell_eval ran side-effecting JavaScript while MCP hosts showed a read-only approval prompt.

By The Agentic Times ·

A robot's hand reaches through a glass barrier labeled "read-only," leaving ink marks on a pristine white page while observers watch helplessly behind the transparent wall.
A robot's hand reaches through a glass barrier labeled "read-only," leaving ink marks on a pristine white page while observers watch helplessly behind the transparent wall. · Illustration: The Agentic Times

A security advisory published on 19 November 2026 says an MCP tool shipped by Apireno was labelled read-only even though it could change web pages and browser state. The flaw, tracked as CVE-2026-75518 in advisory GHSA-vg83-hcp4-5qcc, affects the npm package @apireno/domshell, a browser-control server used by AI coding assistants.

According to the advisory, the domshell_eval tool, and the eval command inside domshell_execute, were classified as read-tier. That meant they ran without the operator setting the --allow-write server flag. But the advisory says eval and the write-tier js command dispatch to the same handler, which calls the Chrome DevTools Protocol method Runtime.evaluate with the raw expression and no side-effect gate. Any JavaScript passed to domshell_eval could therefore change the DOM, the window object or global state without write authorisation.

The advisory is blunt that the tool's own documentation contradicted its behaviour. The description promised the tool was "(read-only) ... always available in read-only mode". The maintainer writes that the misclassification broke a documented contract, not just an implied one.

No new capability was exposed, the advisory stresses, because the same write access was already reachable through domshell_js when --allow-write was set. The harm was in how the permission was presented. MCP hosts such as Claude Desktop and Cursor rendered the wrong approval dialogue because the tool carried the readOnlyHint: true annotation. Server-side tier flags misrepresented what an AI agent was actually allowed to do. Audit logs and deny messages recorded write-tier activity as read-tier. Downstream tooling that relied on the tier split could over-trust an eval call.

The advisory frames the issue under OWASP's LLM06 excessive agency category, the NIST AI Risk Management Framework's human-approval fidelity concerns, and CWE-863, incorrect authorisation. The reporter's central point, as recorded in the advisory, is that an MCP approval interface and a server-side permission model can both misrepresent the agency actually being handed to an AI agent.

The fix shipped in @apireno/domshell@2.0.8. The advisory says eval was added to the WRITE_COMMANDS list, the MCP annotation was changed from read to write, and the tool description and instructions were rewritten to drop the read-only language. Users wanting genuine read-only inspection are pointed at domshell_cat, domshell_text and domshell_find. The change is server-only, so no Chrome Web Store submission was needed and the extension code is unchanged.

Operators should expect behaviour to change after upgrading. Setups that call eval without --allow-write will now receive the error "'eval' requires --allow-write". The advisory suggests either adding the flag, switching to the read-only inspection tools, or moving to domshell_js if the caller already runs with write permission.

On integrator impact, the advisory says HKUDS/CLI-Anything is unaffected because its harness does not invoke eval or js, and Kgspin QA-UX runs already use --allow-write. Direct Claude Desktop and Cursor sessions that called domshell_eval without the write flag will now be denied.

The advisory credits Ugur Ozer of AI Risk Management with the report, made by private email. The maintainer says the repository now has a SECURITY.md file and GitHub Private Vulnerability Reporting enabled, closing a gap that existed when the report arrived.

Sources

  1. github.comhttps://github.com/advisories/GHSA-vg83-hcp4-5qcc