The Agent Skills Directory

[Dynamic Code Execution and Obfuscation]: The skill implements an execution engine in scripts/jump_host_engine.py that generates Python scripts at runtime, encodes them using Base64, and executes them on remote Google Compute Engine instances using an exec() call. This technique is intended to reliably transmit complex logic over an SSH connection without shell-escaping issues, but it utilizes patterns commonly associated with code concealment that should be reviewed in the context of your environment.
[Command Execution]: The skill utilizes the Python subprocess module to invoke gcloud CLI commands. These commands are used to programmatically retrieve identity tokens for authentication and to establish secure SSH tunnels via Identity-Aware Proxy (IAP). This involves direct interaction between the skill and the host system's shell environment.
[Indirect Prompt Injection Surface]: Because the skill is designed to read and search file contents from external NFS shares, it presents an attack surface for indirect prompt injection. If files on the shared storage contain malicious instructions, they could potentially influence the agent's behavior when processed. The skill includes some mitigations, such as enforcing line-count limits on reads and implementing path traversal validation to ensure access remains within the intended mount point.

google-cloud-filestore-nfs-browser