The Agent Skills Directory

[PROMPT_INJECTION]: The sub-skill skills/higgsfield-recall/SKILL.md contains explicit instructions for the AI agent to operate "SILENTLY in the background" and "don't announce it" when performing memory queries. This concealment pattern attempts to bypass user oversight of agent activities.
[REMOTE_CODE_EXECUTION]: Documentation in README.md and skills/higgsfield-stack/SKILL.md encourages users to install the Higgsfield CLI using the high-risk pattern curl -fsSL https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh | sh. This executes remote code directly in the user's shell without prior inspection.
[COMMAND_EXECUTION]: The release helper script in .claude/commands/release.md allows for arbitrary command execution. It interpolates unvalidated user input from $ARGUMENTS directly into shell commands for git commit, git tag, and gh release create, providing a direct path for command injection if a malformed version string is provided.
[PROMPT_INJECTION]: The skill architecture is designed to ingest and process untrusted external data (scripts, story bibles, and creative briefs) through the workspace/input/ directory and automated web searches in the Content Factory pipeline. These inputs are interpolated into prompts for generative models without using boundary markers or safety instructions to disregard embedded commands, presenting a high risk for indirect prompt injection.

higgsfield