Report vulnerabilities through GitHub private vulnerability reporting. Do not open a public issue. I aim to acknowledge within a week.

These skills are instructions and scripts that an agent executes on your machine, usually with your shell, your gh token, and your repository. Treat anything that widens that blast radius as a vulnerability:

• a script that reads or transmits credentials, tokens, or files outside the working directory;
• a path traversal or symlink escape out of the skill or repo root;
• an unquoted interpolation that lets repository content, an issue body, or a filename execute as a command;
• prompt injection that steers an agent into destructive or exfiltrating behavior;
• committed secrets, or private data left in a bundled asset.

• An agent doing something you told it to do, including a destructive command you approved.
• A skill that needs a dependency you have not installed.
• Anything in an upstream tool a skill shells out to (gh, agent-browser, ImageMagick). Report those to their own maintainers.

• Read a skill before you install it. Every one is plain Markdown and Node.
• ui-evidence and to-spec/to-tickets publish to GitHub with your token. Both stop for confirmation first. Keep it that way.
• Screenshots, GIFs, and specs are permanent once published. Redact before upload, not after.