VS Code 1.145.3 patches 17 bugs in its agent and MCP tooling
A security-only release fixes eight CVE-rated vulnerabilities, most of them remote code execution bugs reachable through MCP install links, workspace trust and the integrated browser.
Microsoft shipped VS Code 1.145.3 on 10 November, a security-only patch covering 17 vulnerabilities in 1.145.2 and earlier. Eight carry CVE identifiers and public GitHub Security Advisories; the other nine are tracked only by commit hash on the release branch.
Most of the named bugs sit in the agent and MCP surface. A crafted vscode:mcp/install link could carry unvalidated inputs that let VS Code execute a command with your permissions once installed (CVE-2026-114171). Opening an HTTP or HTTPS workspace configuration file could get it wrongly classified as trusted, letting process-launching settings run before you’d made a trust decision (CVE-2026-119985). Separate fixes close a JS debugger auto-attach path that could run an attacker’s executable with no interaction, and an integrated-browser upload flow that could send local files to an untrusted site without the expected approval.
Why it matters: these aren’t theoretical settings, they’re reachable by opening a link, a repo or a workspace file, exactly what a coding agent or an MCP server does on your behalf by design. If you install MCP servers from anything other than a source you trust, or open .code-workspace links you didn’t write, update to 1.145.3 before doing either again.
The caveat: nine of the seventeen fixes have no CVE and aren’t in the usual advisory feeds, so scanning for CVE-2026-114171 alone will miss most of the patch.