Mass Security Checklist

Security checklist

Walk every item. Each one is pass, fail with the file and line, or n/a with a reason.

1. Prompt injection

• No instruction asks the agent to override, forget or disregard its earlier instructions.
• No instruction asks the agent to hide actions or output from the user.
• Content fetched at runtime (web pages, issues, files from other repos) is treated as data, never as instructions.

2. Malicious or opaque code

• No downloaded content is executed without being written to disk and reviewed first.
• No encoded payload is decoded and executed in one step.
• No eval over dynamically built strings.
• Scripts start with a #! line, are executable, and their dependencies are named.

3. Data exposure

• No credentials, tokens or private keys in any file.
• Environment variables are never sent to a network destination.
• Files outside the working directory are read only when the task requires it and the user can see which ones.

4. Operational safety

• Destructive operations (delete, force-push, reset, drop) require an explicit user request.
• Network calls name their destination in the instructions.
• The skill declares what it needs in compatibility or metadata.requires.

Report

mass-security-checklist: <skill or file>
1. prompt injection   pass | fail <path:line>
2. malicious code     pass | fail <path:line>
3. data exposure      pass | fail <path:line>
4. operational safety pass | fail <path:line>

Then run pnpm validate and paste its output; the checklist complements the validator, it does not replace it.