The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted code from Unity projects through tools like unity_read_script. This data ingestion surface is vulnerable to indirect prompt injection if project files contain malicious instructions. The skill lacks boundary markers or delimiters to differentiate between code data and agent instructions. While the agent has the capability to read files and write reports, there is no sanitization of the input scripts before they are processed by the model.

Unity Review