The Agent Skills Directory

EXTERNAL_DOWNLOADS (MEDIUM): The skill instructs the agent to clone a repository from an untrusted GitHub source (https://github.com/Eng0AI/magic-portfolio-template.git). This source is not part of the trusted organizations list and its contents cannot be verified statically.
COMMAND_EXECUTION (MEDIUM): After cloning the external repository, the skill executes npm install, npm run build, and npm run dev. These commands trigger the execution of scripts (like preinstall or postinstall) and binaries defined within the untrusted repository, creating a path for remote code execution.
CREDENTIALS_UNSAFE (LOW): The deployment section uses environment variables like $VERCEL_TOKEN. While standard for deployment, there is a risk that malicious code in the cloned repository could attempt to exfiltrate these credentials during the build or deploy process.
INDIRECT_PROMPT_INJECTION (LOW): The skill processes external data (the cloned repository).
Ingestion points: git clone of the external repository.
Boundary markers: None present; the agent treats the cloned files as legitimate project structure.
Capability inventory: File system access, network access via npm, and subprocess execution via npm run.
Sanitization: None; the agent is instructed to move and execute the files directly.

magic-portfolio