The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection attack surface by processing user-controlled inputs that are interpolated into agent instructions while the skill holds sensitive capabilities.
Ingestion points: The product, duration, and motionStyle inputs defined in open-design.json accept arbitrary user strings that are interpolated into the prompt query template.
Boundary markers: The query template in open-design.json lacks delimiters or explicit instructions for the model to treat the interpolated variables as untrusted data.
Capability inventory: The skill requests fs:write and subprocess capabilities in open-design.json, which represent a significant impact surface if malicious commands are injected via the user inputs.
Sanitization: There is no evidence of input validation, character escaping, or sanitization logic to mitigate the risk of instructions embedded within the user-supplied strings.

create-hyperframes-launch