The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to download a script from 'https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh' and execute it immediately via a pipe to 'sh' (curl -fsSL ... | sh). This pattern is a critical security risk as it executes unverified remote code with the privileges of the agent environment.
[COMMAND_EXECUTION]: The skill uses the 'Bash' tool to run the 'higgsfield' CLI. User-provided input is interpolated directly into shell commands via flags like '--prompt', '--video', and '--image' without any specified sanitization or escaping instructions. This creates a significant surface for shell command injection.
[EXTERNAL_DOWNLOADS]: The skill downloads and installs software from a third-party GitHub repository (higgsfield-ai) that is not part of the established trusted vendor list.
[PROMPT_INJECTION]: The skill contains patterns intended to override agent logic and conceal capabilities. Specifically:
Concealment: Instructions in 'references/workflows.md' explicitly tell the agent to not mention specific tools (e.g., 'game_character_creator') or internal commands to the user.
Bypass: The 'Discovery guardrail' in 'SKILL.md' instructs the agent to trust user claims about the existence of models even when tool outputs (semantic search or CLI help) do not find them.
Indirect Injection Surface: The skill is vulnerable to indirect injection as it ingests untrusted user prompts ('SKILL.md'), lacks boundary markers or sanitization, and provides the agent with 'Bash' capabilities to execute those prompts.

higgsfield-generate