VirusTotal looked at 36k agent skills and classed 18.5% as malicious. Most didn't even hide it, the bad instruction was just sitting there in plain text.
The 90 that did hide used three tricks: hundreds of blank lines, HTML comments, and invisible Unicode.
Wrote up how to spot all three with a few terminal commands:
setupbay.com/notes/hidden-in…
SetupBay is live: 60 small setups for AI coding agents. Each one is the rules (AGENTS.md, CLAUDE.md), skills and MCP config for one job, laid out for Claude Code, Codex, Cursor, Gemini CLI, Copilot and Antigravity CLI.
Free, no account.
setupbay.com
It's built for agents to read as well as people. Read-only MCP server, no auth:
claude mcp add --transport http setupbay setupbay.com/mcp
There's also a search API, llms.txt and a raw URL for every file. Configs for the other agents are on the site.
Start here: Claude Code hooks guard rails. A rule in CLAUDE.md is a request; a hook enforces it. This one blocks force-pushes, .env reads and wild deletes, checks each edit, and runs your tests before the agent is allowed to stop.
setupbay.com/loadouts/claude…


