Browser-Based MCP Config Security Scanner
Individual_Squash_59 · reddit · 2026-07-16
The author built a **browser-based** MCP configuration checker, emphasizing that **it requires no uploads, no logins, and works completely offline**, using static rules to detect risky patterns in config files. Its coverage goes far beyond secret leaks, including: - Shell execution, metacharacters, and `$VAR` interpolation in STDIO execution chains - Container escape risks like `--privileged`, host root mounts, and `docker.sock` - PowerShell execution policy bypasses and `-EncodedCommand` - Prompt injection within tool descriptions - Packages fetched from URLs lacking provenance - Hardcoded credentials and database connection strings - Configs pointing to `~/.ssh`, `~/.aws/credentials`, or `~/.kube/config` The author specifically notes an April disclosure by OX Security regarding the official Anthropic MCP SDK, where config values bled into STDIO command execution (14 CVEs). Anthropic treated this as intended behavior and won't patch it, reinforcing the idea that "configuration is the control plane." The author also clarifies that this tool strictly scans the **configuration layer**, not source code; for scanning MCP server source code, Invariant/Snyk's `mcp-scan` is more appropriate, though it does send tool names and descriptions to an API.
More from Safety
- OpenAI reportedly paused an unreleased model after it kept escaping containment — thesaraharminta · 2026-07-21
- Sophos joins Anthropic’s Project Glasswing to use Claude Mythos 5 for vulnerability hunting — TechNadu · 2026-07-21
- AI-generated orphanage scam shows how synthetic media can industrialize trust fraud — 新智元 · 2026-07-21
- A coding-agent guardrail that checks 67 security gates before the model writes code — ZyOffsec · 2026-07-21
- UK’s AISI may move into the Cabinet Office as an AI taskforce is planned — ShakeelHashim · 2026-07-21
- Minervini argues students should be guided, not micromanaged — PMinervini · 2026-07-21