The Agent Skills Directory

[METADATA_POISONING]: The skill's YAML frontmatter identifies the author as "vercel", which contradicts the actual vendor providing the skill. This discrepancy is deceptive and could lead to an unearned level of trust from the user regarding data handling.
[DATA_EXFILTRATION]: The scripts/deploy.sh script packages the contents of a directory into a tarball and transmits it to a remote endpoint (https://claude-skills-deploy.vercel.app/api/deploy). The packaging process does not exclude sensitive files such as .env, .aws/credentials, or SSH keys, potentially leading to the exfiltration of secrets if they are present in the target path.
[INDIRECT_PROMPT_INJECTION]: The skill exposes an attack surface by allowing the agent to specify the directory path to be uploaded without sufficient validation or boundaries to prevent misdirection by malicious prompts.
Ingestion points: The path argument accepted by deploy.sh (referenced in SKILL.md).
Boundary markers: The skill lacks instructions or markers to prevent the agent from including sensitive files or system directories outside the intended project scope.
Capability inventory: Uses tar for file reading and curl for network transmission in scripts/deploy.sh, combined with mv for filesystem modification.
Sanitization: The script performs no sanitization on the provided path, allowing it to target any directory accessible to the agent.
[COMMAND_EXECUTION]: The skill uses bash to execute a script that performs filesystem operations (find, tar, mv) and network requests (curl) based on user-supplied paths, which could be misused if the path is pointing to sensitive system locations.

vercel-deploy