The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill demonstrates the execution of agents directly from remote Git repositories using the -r flag (e.g., gitagent run -r https://github.com/user/agent). This allows the agent to download and run logic from arbitrary third-party sources, which may contain malicious code or unintended instructions.
[COMMAND_EXECUTION]: The skill utilizes a CLI tool (gitagent) to perform system-level operations, including cloning remote repositories to a local cache at ~/.gitagent/cache/ and launching processes for various agent adapters.
[EXTERNAL_DOWNLOADS]: The skill promotes downloading content from external Git repositories. While the examples use placeholders, the functionality encourages the retrieval of unverified data from the internet.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted data from remote repositories and local framework files. Ingestion points: Remote repository contents, .cursorrules, crew.yaml, and .gitagent_adapter files. Boundary markers: The skill does not define specific delimiters or instructions to prevent the agent from obeying embedded instructions in these files. Capability inventory: The tool can clone repositories, write to a local cache, and execute agent logic. Sanitization: There is no evidence of content sanitization or safety checks performed on the downloaded repository data before it influences the agent's behavior.

run-agent