The Agent Skills Directory

[COMMAND_EXECUTION]: The skill directs the agent to execute local Python scripts using shell commands that incorporate unsanitized user input. Specifically, the command python3 scripts/seedance_lint.py --preflight --model seedance_2_0 "<prompt text>" passes user-provided text directly into a shell-executed string. This creates a severe command injection vulnerability where an attacker could provide input containing shell operators (like ;, &&, or backticks) to execute arbitrary code on the host system.\n- [PROMPT_INJECTION]: The skill includes sections titled "The Filter Model" and "The Rewrite Playbook" that are explicitly designed to teach how to circumvent AI content filters. These sections provide concrete examples of how to reword prohibited content—including real-world individuals, copyrighted brand names, and violent actions—specifically to evade detection by safety models.\n- [COMMAND_EXECUTION]: Additional instructions utilize shell commands to update memory databases (e.g., python3 scripts/higgsfield_memory.py update-filter <id> ...), which similarly rely on interpolating user-provided identifiers and outcomes into command line arguments without clear validation or sanitization.

higgsfield-seedance