CVE-2026-62964, CVE-2026-75612, CVE-2026-75518: Three MCP servers let AI agents break out of tool limits

Three MCP servers that expose tools to AI agents shipped with path-traversal or permission-bypass flaws that let a prompt-injected agent or malicious client break out of the intended sandbox.

Christopher Ort
By Christopher Ort · 5 min read
Share
A developer's workstation with multiple terminal windows open, code editor in the background, and a hand reaching for the keyboard, photographed from above in a modern office setting.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. What you should do
  5. What it means
  6. Sources

NVD published CVE entries on 18 November 2026 for three MCP (Model Context Protocol) servers with vulnerabilities that let AI agents or malicious clients escape the intended boundaries of their tool calls. The flaws affect Gortex (a code-intelligence engine), yt-mcp (a YouTrack integration for Claude Code), and DOMShell (a Chrome automation server). All three have patches available.

i10X Quick Take

Quick Take

  1. What: Three MCP servers (Gortex, yt-mcp, DOMShell) shipped with path-traversal or permission-bypass bugs that let tool calls write files, read secrets, or mutate state outside the intended scope.
  2. Who is affected: Developers running Gortex before 0.55.0, yt-mcp 1.18.0 and earlier, or DOMShell before 2.0.8 with AI agents or untrusted MCP clients.
  3. CVEs: CVE-2026-62964 (Gortex, CVSS 7.8 high), CVE-2026-75612 (yt-mcp, CVSS 7.5 high), CVE-2026-75518 (DOMShell, CVSS 4.8 medium).
  4. Patches: Gortex 0.55.0, yt-mcp 1.18.1, DOMShell 2.0.8.
  5. What to do: Upgrade to the patched version if you use any of these tools; for yt-mcp, avoid the file_path mode of add_attachment until upgrading.

In plain English

MCP servers are add-ons that let AI agents (like Claude or coding assistants) do things such as reading files, editing code or driving a web browser. Three of them had bugs that let an agent do more than it should if it was tricked by hidden instructions or controlled by a malicious app. Gortex could write files anywhere on the computer (which can lead to running attacker code), yt-mcp could read any file and upload it to a YouTrack ticket, and DOMShell could change web pages while its approval screen called the action read-only. All three projects have released fixes.

What you should do: If you use one of these tools, update it; if not, you do not need to do anything.

What happened

The three vulnerabilities share a common pattern: an AI agent or MCP client can supply arguments to a tool call that escape the intended confinement or permission boundary.

Gortex (CVE-2026-62964)

Gortex is a code-intelligence engine for AI agents and IDEs. The write_file, edit_file, and move_inline tools in versions before 0.55.0 accepted absolute file paths without checking that they fell within the indexed repository roots. The resolveFilePath function exempted absolute paths from confinement, and the write handlers never called the guardSymlinkWithinRepo check that the read handlers used. A caller who could influence MCP tool arguments (a prompt-injected agent or a malicious client) could write any file the daemon user had access to, including startup scripts (~/.zshrc, ~/.bashrc), Git hooks (.git/hooks/pre-commit), or systemd units, leading to arbitrary code execution on the next trigger. The flaw violated Gortex’s own documented security invariant that file-path resolution is confined to indexed repository roots. The issue is fixed in version 0.55.0.

yt-mcp (CVE-2026-75612)

yt-mcp is a YouTrack MCP server for Claude Code, distributed as a pip package. The add_attachment tool in versions up to and including 1.18.0 accepted a caller-controlled file_path, checked only that the file existed, read it without directory confinement, and uploaded the bytes to a YouTrack issue. A compromised MCP client or an LLM steered by indirect prompt injection could read any file accessible to the server process (system password files, SSH private keys, environment or credential files) and exfiltrate the contents into a YouTrack issue the caller could access. A related, lower-severity traversal in get_project_health allowed a caller-controlled project value to be interpolated into snapshot filenames, permitting parent-directory traversal to write JSON files outside the snapshots directory. The issue is fixed in version 1.18.1. Until upgrading, operators can avoid the file_path mode of add_attachment and use the content or content_base64 mode instead, where the file is read in the client’s own trust context. The fix disables the file_path mode by default and only permits it for directories the operator explicitly allowlists via YOUTRACK_ATTACHMENT_ROOTS, resolved with realpath and commonpath (symlink-safe, sibling-prefix-safe), with a size cap.

DOMShell (CVE-2026-75518)

DOMShell is a Chrome automation project whose MCP server exposes Unix-style tools for AI agents, distributed as the npm package @apireno/domshell. In versions before 2.0.8, the domshell_eval tool and the eval command inside domshell_execute were classified as read-tier operations that could run without the --allow-write server flag. However, eval and the write-tier js command both dispatched to the same handler in the extension and called CDP Runtime.evaluate with the raw expression string and no side-effect gate. Any JavaScript expression could therefore mutate DOM, window, or global state through domshell_eval without the operator’s --allow-write authorization. The tool description explicitly described domshell_eval as read-only and always available in read-only mode, so the misclassification contradicted the documented contract. MCP hosts (Claude Desktop, Cursor, etc.) rendered the wrong approval UI (readOnlyHint: true), server-side tier flags misrepresented what an AI agent was actually authorized to do, and audit logs and deny messages labeled write-tier activity as read-tier. The issue is fixed in version 2.0.8, which adds eval to the write-tier command list, changes the domshell_eval MCP annotation from read to write, and updates the tool description to remove read-only language. Callers that invoke eval without --allow-write will now receive an error; options include adding --allow-write if the write capability is needed, migrating read-only inspection to domshell_cat, domshell_text, or domshell_find, or migrating to domshell_js if the caller already runs with --allow-write.

What you should do

  • Gortex users: Upgrade to version 0.55.0.
  • yt-mcp users: Upgrade to version 1.18.1. Until upgrading, avoid the file_path mode of add_attachment; use content or content_base64 instead.
  • DOMShell users: Upgrade to version 2.0.8 (npm install @apireno/domshell@2.0.8). If you call eval without --allow-write, you will need to add that flag or migrate to read-only tools.

What it means

MCP servers are a new attack surface in AI-assisted development workflows. These three vulnerabilities show that tool-call boundaries (file-path confinement, read vs. write permissions) must be enforced consistently across all code paths, not just the read side. Prompt injection and malicious clients are the standard MCP threat model: an attacker who can steer an agent’s tool arguments can exploit path-traversal or permission-bypass bugs to read secrets, write files, or execute code on the host running the MCP server. The fixes apply the same confinement and permission checks that were already present on the read path to the write path, and correct the tier classification so that approval UIs, server flags, and audit logs accurately represent the agency being granted to an AI agent.

Path and permission bugs in MCP tools are a recurring pattern: see our reports on the path traversal flaw in Code Graph RAG MCP Server and on path traversal in mcp-atlassian. New cases are tracked in our AI Security hub.

Sources

  • NVD CVE-2026-62964
  • GitHub Advisory GHSA-w42c-h7hr-f67p (Gortex)
  • NVD CVE-2026-75612
  • GitHub Advisory GHSA-99mq-fjjc-6v9j (yt-mcp)
  • NVD CVE-2026-75518
  • GitHub Advisory GHSA-vg83-hcp4-5qcc (DOMShell)
Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Editor's picks

  1. CVE-2026-133161: Pydantic AI Harness ANSI stripping flaw blocks event loop
  2. Circus Brings Its CA-1 Robot Kitchen Into a Hospital, Serving Fresh Meals Around the Clock at Schoen Clinic Munich
  3. Metano Emerges From Stealth With a $12 Million Seed to Secure AI Agents Across Endpoints, Cloud and SaaS
  4. IEA World Energy Outlook 2026: AI Boom Doubles Data Center Electricity Use to 1,000 TWh by 2030
  5. Know Your Agent: Singapore’s MAS Plans a Protocol for AI Agents in Banking and Warns Users Not to Share Their Logins