Browser-automation MCP servers broke their read-only modes: Safari MCP read-only tools could run arbitrary page JavaScript via an escaping bug (fixed 2.22.15), and DOMShell's domshell_eval ran writes without --allow-write (CVE-2026-75518, fixed 2.0.8)
Two MCP servers that let AI agents drive a real browser promised a read-only tier that didn't hold. Safari MCP (npm safari-mcp) put tool arguments inside single-quoted JavaScript strings in 16 places and escaped the quote but not backslashes, so an argument with a backslash and a quote broke out and ran as page JavaScript. Seven of those arguments belong to read-only tools exempt from the tab-ownership check, so a client limited to read-only tools (safari_evaluate and safari_run_script excluded) could still run arbitrary JavaScript, even in tabs the session doesn't own (fixed 2.22.15). DOMShell (npm @apireno/domshell) classified domshell_eval as read-tier although it shares the Runtime.evaluate path with the write-tier js command, so it could mutate the page without --allow-write while MCP hosts showed a read-only approval UI (CVE-2026-75518, fixed 2.0.8).
What it is
Safari MCP (GHSA-4rj4-rcrq-gr92, no CVE): affected read-only arguments are filter in safari_network_details and safari_extract_links, level in safari_console_filter, selector in safari_snapshot, safari_accessibility_snapshot and safari_extract_tables, and dbName/storeName in safari_get_indexed_db. Other affected arguments include ref in safari_select_option, the key in safari_press_key and file names in safari_upload_file / safari_paste_image. The fix routes every site through a shared escJsSingleQuote helper. Affected: safari-mcp < 2.22.15 (released 10 October 2026); the advisory reached the GitHub Advisory Database on 19 November 2026. There is no configuration workaround.
DOMShell (GHSA-vg83-hcp4-5qcc, CVE-2026-75518): the tool advertised readOnlyHint: true, so hosts such as Claude Desktop and Cursor rendered the wrong approval UI, and audit logs labelled write activity as read. It exposes nothing beyond what domshell_js already allows with --allow-write. Affected: @apireno/domshell < 2.0.8; NVD published the CVE on 18 November 2026.
In both cases, text the agent reads from a web page or other untrusted source can steer the arguments it passes, so a 'read-only' allowlist was not a real boundary.
How to check
- npm ls -g safari-mcp @apireno/domshell, or check the version pinned in your MCP client config (npx/uvx args).
- Check whether you relied on read-only tool allowlists or on running DOMShell without --allow-write.
How to fix
- Upgrade safari-mcp to 2.22.15 or later and @apireno/domshell (MCP) to 2.0.8 or later.
- Use a separate browser profile for agent automation, without your logged-in sessions.
npm ls -g safari-mcp @apireno/domshell 2>/dev/null
grep -rnE 'safari-mcp|domshell' ~/.cursor/mcp.json ~/Library/Application\ Support/Claude/claude_desktop_config.json 2>/dev/null Sources
- GitHub advisory GHSA-4rj4-rcrq-gr92: Safari MCP JavaScript injection through single-quoted tool arguments (backslash not escaped) ↗
- Safari MCP v2.22.15 release ↗
- GitHub advisory GHSA-vg83-hcp4-5qcc: domshell_eval classified as read-tier can execute side-effecting JavaScript without --allow-write ↗
- NVD: CVE-2026-75518 ↗
- criticalMalicious npm package @guicai777/mcp-doctor posed as an MCP diagnostic tool and dropped a hidden Windows executable, with a README telling users to run mshta (MAL-2026-20295)
- criticalCVE-2026-77359: Jupyter MCP Server's /api/connect and /api/stop routes skipped the token and Origin checks, so a website or network caller could repoint it to an attacker's Jupyter backend (fixed 1.0.3)
- criticalMalicious npm packages posing as AI SDKs and MCP servers: character-ai-sdk, char-ai-sdk, @mibbp/agent-observer, mrnima-openai-chatgpt-bot and @fdcasj_dfew/md-doc-mcp
- criticalCVE-2026-76807: Doist's official Todoist MCP HTTP server let any website use DNS rebinding to read, create and delete your Todoist data (fixed 10.3.3)