The Agent Skills Directory

Dynamic Code Assembly and Execution: The scripts/jump_host_engine.py component constructs Python scripts at runtime and executes them on remote Compute Engine instances. It uses a pattern where the script is base64-encoded and passed to a remote shell via python3 -c "exec(base64.b64decode(...))". While this technique is employed here to ensure reliable script transmission over SSH without shell-escaping issues, it is a pattern that can be used to obscure execution logic.
Indirect Prompt Injection Surface: The skill is designed to ingest and display directory trees and file contents from remote NFS shares. Ingesting untrusted data (such as filenames or file content controlled by an external actor) into the agent's context window presents a surface for indirect prompt injection. The skill implements basic safeguards like line limits, entry caps, and binary file detection to mitigate context overflow.
Remote Code Execution Capability: A core feature of the skill is the ability to run code on remote Google Cloud infrastructure (Jump Hosts). This functional capability requires careful management of IAM permissions, such as roles/iap.tunnelResourceAccessor and OS Login roles, to prevent unauthorized access to the execution environment.
System Command Execution: The CLI tool uses the subprocess module to invoke gcloud commands for identity token retrieval and SSH tunneling. This is standard behavior for managing Google Cloud resources but represents a dependency on the local environment's CLI configuration and permissions.

google-cloud-filestore-nfs-browser