The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill executes a remote installation script from claude.ai (a well-known service) using curl | sh on the remote sandbox to prepare the environment.
[COMMAND_EXECUTION]: The offload.sh script uses the eval command to construct and execute shell commands. It interpolates the user-provided TASK_PROMPT directly into a single-quoted string passed to a remote shell. This creates a command injection vulnerability where a prompt containing unescaped single quotes could execute arbitrary commands on the remote sandbox.
[DATA_EXFILTRATION]: The skill automatically collects local environment variables (including sensitive keys like ANTHROPIC_API_KEY) and the .claude/ session history directory, then transfers them to the remote sandbox via rsync. This is intended for task continuity but results in the exposure of local secrets to a remote network environment.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of the @getcompanion/cli and @anthropic-ai/claude-code packages from the NPM registry. These are vendor-owned or well-known packages required for the skill's primary function.

companion-offload