The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill architecture is vulnerable to instructions embedded within the data it processes.
Ingestion points: Untrusted data is ingested from user-provided $ARGUMENTS and the contents of the .guidelines/blade-components.md file.
Boundary markers: The prompt does not utilize delimiters or specific instructions to isolate user-supplied content from agent instructions.
Capability inventory: The skill is granted access to potentially dangerous tools including Bash, Write, and Edit.
Sanitization: No input validation or sanitization routines are specified for handling the ingested external data.

moonshine-components