The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to process external image files and use commands like magick identify -verbose to read image metadata and properties. Malicious files could contain instructions designed to influence agent behavior when the metadata is read.
Ingestion points: Local image files provided or requested by the user for processing (SKILL.md).
Boundary markers: Absent; there are no specific instructions to the agent to disregard or sanitize instructions that might be embedded in image metadata fields like EXIF, IPTC, or XMP.
Capability inventory: The skill utilizes the Bash, Read, Grep, and Glob tools to execute commands and inspect file contents.
Sanitization: Absent; the documentation does not include steps to filter or escape content retrieved from image metadata.
[COMMAND_EXECUTION]: The skill heavily relies on the Bash tool to execute magick and mogrify commands for image manipulation.
Evidence: Multiple code examples in SKILL.md demonstrate shell command patterns using wildcards (e.g., *.jpg) and shell variables (e.g., "$img"). While the examples use quotes to mitigate simple word splitting, the reliance on shell execution creates a surface for command injection if the agent constructs commands using unsanitized user-provided filenames containing shell metacharacters.

imagemagick-conversion