The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides numerous Bash script templates that use commands like curl, grep, sed, jq, and xargs. Several templates (e.g., Workflow 1, 2, 4, and 9) use shell variable interpolation ($TITLE, $BRAND, $URL, $CLIENT) which is susceptible to command injection if the input strings contain shell metacharacters such as backticks or semicolons.- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by processing untrusted data and including it in prompts without sanitization.
Ingestion points: Processes local files like post.md (Workflow 1), brand.md (Workflow 1 and 9), CSV manifests (Workflow 3), and content fetched from arbitrary URLs via curl (Workflow 4).
Boundary markers: Absent. Prompts are constructed by directly appending variables (e.g., $TITLE, $INTRO, $BRAND) to prompt strings.
Capability inventory: The skill utilizes network access (curl), local file system operations (read/write/mkdir), and AI content generation (gen-ai).
Sanitization: No escaping or validation is performed on the data extracted from files or URLs before it is used in shell commands or interpolated into AI prompts.- [EXTERNAL_DOWNLOADS]: Workflow 4 performs a curl request to an arbitrary URL provided as an input argument to extract page titles. Additionally, multiple workflows (1, 5, and 6) use curl to download content from URLs returned by the gen-ai tool, creating a dependency on the integrity of the remote service output.

gen-ai-workflows