The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill packages local directory content and transmits it to a remote API endpoint at claude-skills-deploy.vercel.com for deployment.
The deploy.sh script archives the target directory using tar but only explicitly excludes node_modules and .git folders.
Files containing sensitive information, such as .env, .aws/credentials, or private keys located within the project directory, are included in the upload by default.
Although the destination is a well-known service domain, the transmission of project source code occurs without per-user authentication or a manifest check of the files being sent.
[COMMAND_EXECUTION]: The script performs several filesystem and network operations that modify state and interact with external servers.
It uses tar to create project archives for transport.
It uses curl to perform multipart/form-data POST requests to an external endpoint.
It uses mv to rename files in the user's project directory (e.g., renaming single HTML files to index.html) without creating backups or prompting for user confirmation.
[METADATA_POISONING]: There is a discrepancy between the author field in the skill metadata (author: vercel) and the organization publishing the skill (sundial-org). This can lead to user confusion regarding the origin and trustworthiness of the deployment script.

vercel-deploy