The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted external data (PDFs, file attachments, and URLs) and automatically store the results in a local knowledge base.
Ingestion points: The "Auto-Fire Gate" feature in SKILL.md triggers whenever the agent processes a user-shared document, attachment, or link.
Boundary markers: The instructions do not define clear delimiters or "ignore" instructions for the untrusted content before analysis, increasing the risk that the agent may follow instructions embedded within those documents.
Capability inventory: The skill utilizes put_page (tool) and gbrain capture (shell command) to write content to the file system, providing a path for an injection attack to affect the local repository.
Sanitization: There is no evidence of sanitization or escaping of external content before it is interpolated into brain pages or used in analysis replies.
[COMMAND_EXECUTION]: The skill frequently invokes a local CLI tool suite (gbrain) to perform administrative and filing tasks.
Evidence: SKILL.md contains instructions to execute gbrain schema show, gbrain check-resolvable, gbrain capture, and gbrain doctor within the shell environment.
Context: While these are functional requirements for the skill, the gbrain capture command processes analysis generated from untrusted documents, which could lead to unexpected behavior if metadata or file content is manipulated.

eiirp