Guide

Can a Claude or Codex Skill hide malware? How to check a Skill or MCP server before installing

Short answer

Yes: a Skill, MCP server or package can contain commands that run on your computer, read your .env files and keys, send data to unknown servers or hide instructions that hijack your assistant. Hackaiz reads the files before you install them, without running anything, and returns Safe, Attention, Danger or Incomplete with the exact line as evidence.

By the Hackaiz team · HACKTUALIZ Inc. · Updated

Check a Skill or MCP server

Check in 3 steps

  1. 1Paste the text of a SKILL.md, or choose the files of a Skill, an MCP config (mcp.json), a package.json or a workflow.
  2. 2Hackaiz reads the files without running anything; secret files (.env, keys) are never sent.
  3. 3You get a clear verdict, the exact line as evidence and a safer alternative when one exists.

Safe

None of the listed checks found risky behaviour.

Attention

Review this point before installing.

Danger

Do not install: dangerous behaviour found.

Incomplete

Some files could not be read: never shown as safe.

Each check uses credits based on content size. Also available in VS Code/Cursor, Claude via MCP and the command line.

Check now

What Hackaiz looks for

  • Remote scripts piped to a shell (curl | bash) and hidden system commands.
  • Reading .env files, SSH keys, cloud credentials or environment secrets.
  • Data sent to servers that are not well-known services.
  • Hidden instructions for the AI and invisible characters.
  • Encoded (obfuscated) code, auto-start installation, destructive commands.
  • Install scripts, unpinned versions and names imitating popular packages.
  • MCP configs giving access to the whole disk; GitHub workflows that outsiders can hijack; containers running as root.

Dangerous combinations

Some abilities are acceptable alone but dangerous together. A component that can both read your secrets and send data outside is flagged Danger even when each part looks minor.

Where to check

  • On hackaiz.com: Check a component (paste a SKILL.md or choose files).
  • In VS Code or Cursor: Hackaiz offers a check when a Skill, MCP config, package manifest or workflow is added.
  • In Claude, ChatGPT or Cursor through the Hackaiz MCP connector: the check_component tool, after you approve.
  • In a terminal or CI: hackaiz check <path>, which exits with an error on Danger and can output SARIF.

Honest limits

These are static and reputation checks: nothing is executed, so behaviour that only appears at run time can be missed. Safe means none of the listed checks found risky behaviour; it is not a guarantee or a certification. Files that could not be read make the result Incomplete, never Safe.

Frequently asked questions

Do I still need to read every Skill myself?

Hackaiz does the first reading for you and points to the exact lines that matter. For sensitive projects, review the flagged lines before installing.

Are my files stored?

No. The files are analysed in memory and not kept. Secret files such as .env and private keys are never sent.