The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's instructions in SKILL.md and references/brand.md direct the agent to use curl to download images from external URLs. This practice allows for arbitrary network connections and can be exploited for data exfiltration if the agent is induced to include sensitive information in the URL parameters.
[DYNAMIC_EXECUTION]: The moda canvas edit feature executes batch JavaScript mutations on the canvas. As documented in references/edit-code.md, the sandbox is considered a "speedbump" and not a hard security boundary, presenting a risk of sandbox escape and unauthorized code execution within the host environment.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted content from the web and local files to generate document layouts. It does not provide guidance on sanitization or the use of boundary markers, creating a vulnerability to indirect prompt injection where malicious instructions in the source data could manipulate the agent's behavior.
[EXTERNAL_DOWNLOADS]: The workflow involves downloading assets from arbitrary external domains identified during brand extraction or web research, which increases the risk of interacting with malicious servers.
[DATA_EXFILTRATION]: By combining the ability to read local files (via Read, Glob, Grep) with network-capable instructions like curl, the skill creates a potential pathway for the exfiltration of sensitive local data to external endpoints.

moda-document