The Agent Skills Directory

[COMMAND_EXECUTION]: The infrastructure implementation patterns in references/infrastructure-patterns.md and scripts/sandbox_manager.py demonstrate unsafe command construction. Specifically, user-provided identity fields such as 'name' and 'email' are directly interpolated into shell commands (e.g., git config user.name "{user.name}") using os.system and execute_command. This allows an attacker to execute arbitrary commands within the sandbox by providing a crafted user identity.
[EXTERNAL_DOWNLOADS]: The skill's design involves automated fetching of external code. It includes logic to clone GitHub repositories based on user-supplied URLs and to install dependencies from public registries using npm install and pip install during the sandbox image build process.

hosted-agents