The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using arguments derived from user input or external data, which could lead to command injection if the inputs are not sanitized.
Evidence: md5 -q <input_file> where <input_file> is an argument provided to the skill.
Evidence: rg -il "hash brown" --glob "*.yaml" <DATA> where the search string is determined by the agent after analyzing an image.
[DATA_EXFILTRATION]: The skill facilitates sending local image data or data derived from the file system to an external service (Imgur).
Evidence: The workflow utilizes a social-to-imgur skill to get a "permanent Imgur thumbnail URL" for both social media links and local files resolved to shortcodes.
[PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection because it ingests data from untrusted external sources (URLs) and local files, which are then processed by the agent.
Ingestion points: User-provided recipe URLs (YouTube, Instagram, etc.) and local YAML files in the grouchygiraffe/data directory.
Boundary markers: None mentioned; the instructions do not specify delimiters or instructions to ignore embedded commands in the fetched recipe data.
Capability inventory: Shell command execution (md5, rg), network access (via chained skills), and file writing to the vault.
Sanitization: No explicit sanitization or validation of the fetched recipe content is described before it is processed or written to the vault.
[EXTERNAL_DOWNLOADS]: The skill references and fetches content from well-known media and social services.
Evidence: Targeted domains include youtube.com, instagram.com, and imgur.com.

grouchygiraffe-ingest