The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external data from user-provided URLs and prompt strings to drive video generation models. This architecture presents a potential surface for indirect prompt injection attacks where an attacker-controlled product page or malicious user input could attempt to subvert the agent's behavior.
Ingestion points: Data enters the context via the url parameter in the show_marketing_studio tool and the prompt parameter in generate_video, as documented in SKILL.md sections 3, 6, and 7.
Boundary markers: The instructions recommend using negative constraints (e.g., "ABSOLUTELY NO PACKSHOT" in SKILL.md section 10.8) and structured prompt beats, but lack explicit delimiters to isolate untrusted user content from system instructions.
Capability inventory: The skill uses the generate_video and show_marketing_studio MCP tools. It also references local verification scripts such as scripts/build_index.py and scripts/seedance_lint.py as seen in SKILL.md sections 1 and 14.
Sanitization: There is no evidence of sanitization or escaping mechanisms applied to the external URLs or prompt content before they are processed by the generative models.

higgsfield-marketing-studio