The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions in SKILL.md suggest running a local Python script scripts/seedance_lint.py to validate generated prompts. The command pattern python3 scripts/seedance_lint.py --preflight --model seedance_2_5 "<prompt>" interpolates user-provided content into a shell command. While the input is enclosed in double quotes, this represents a potential command injection surface if the agent does not properly sanitize the <prompt> placeholder before execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted user descriptions to generate complex production instructions. It lacks explicit instructions for the agent to sanitize or delimit user-provided data within its own context, creating a surface for indirect prompt injection. The mandatory evidence chain includes:
Ingestion points: User requests for video generation and scene descriptions entering the agent context as described in SKILL.md.
Boundary markers: Specific brackets like () and {} are defined for the downstream video model, but no boundary markers or instructions are provided to protect the agent's own instruction processing.
Capability inventory: The skill references the capability to execute the seedance_lint.py script via a shell command.
Sanitization: No explicit sanitization, escaping, or validation logic for the external user content is mentioned.

higgsfield-seedance-2-5