AI made everyone a developer. It didn’t make everyone a security engineer.
Vibe coding is fast: clone a repo, paste a key into .env, add an MCP server, tell the agent to run whatever it needs. Each step is a door, and attackers know exactly where they are.
Repo Check: know what a project does before it’s on your Mac.
Anyone can publish a polished open-source project in an afternoon now, and attackers do. An MIT license, a slick README and thousands of stars say nothing about what the code does. A license is legal permission, not a security review.
And the danger isn’t just running the app. A repository can run code the moment you:
- open the folder in VS Code or Cursor;
- install it with
npm installorpip install; - let your AI agent work in it, because files like
CLAUDE.mdand.cursorrulescan carry instructions for your agent, sometimes in text you can’t see.
Paste a GitHub link and Mac Shield reads the project straight from GitHub, in memory, without cloning it or running anything. In seconds you’ll know:
- what runs when you open or install it;
- whether it hides code or instructions for your AI agent;
- whether it reaches for passwords, keys or wallets, or sends data out;
- whether its dependencies are known-malicious;
- who’s behind it: how old the repo and account are, whether its stars came overnight, and what a fork changed from the original.
Already cloned it? Mac Shield checks the folder. Both checks are free.
Pro With Paste Guard on, Mac Shield offers to check a repo the moment you copy its git clone command, and clone checks read new repos in the background as they land.
Audit My Project: a real security review for code you wrote.
Mac Shield runs Cloudflare’s open-source security-audit skill with your own Claude Code. AI agents hunt for flaws attackers could use, and other agents try to disprove each one before it’s reported. You get the findings without the consulting bill.
Your AI agents’ permissions.
“Always allow” is how one poisoned file turns into a disaster. Mac Shield finds the settings that let Claude Code, Codex, Gemini CLI and other agents act without asking, including:
- blanket shell access;
- bypass and “YOLO” modes;
- repos that auto-start their own AI tool servers;
- shell aliases that launch agents with permissions off.
Pro Pro fixes risky agent settings in one click.
MCP servers and poisoned tools.
Every MCP server you add runs with your permissions, and its tool descriptions go straight to your AI agent. Mac Shield lists every server your AI tools start or connect to and flags unpinned packages and broad file access. It can also read each server’s tools for hidden instructions, invisible text and attempts to hijack other tools.
Pro Pro warns you if those tool descriptions quietly change later.
API keys in plain text.
Keys for Anthropic, OpenAI, GitHub, AWS, Stripe, Supabase and more, found in shell profiles, .env files, credential files, AI tool configs, and commands your agent was allowed to run. Values are never shown, kept or sent. You get a link to replace each key.
Dev servers open to your network.
That local server you started for testing may be reachable by everyone on the coffee-shop Wi-Fi. Mac Shield shows which ones, and how to keep them to your Mac.
Flagged packages.
Every lockfile in your projects is checked against OSV for known-malicious and vulnerable packages, grouped by project, with the command to update them.
