The Agent Skills Directory

[PROMPT_INJECTION]: The 'Data Processing Pipeline' section provides a script that reads external file content (cat $file) and interpolates it directly into an LLM prompt for the belt CLI. This pattern lack markers to distinguish data from instructions, creating an indirect prompt injection surface.
Ingestion points: SKILL.md (Data Processing Pipeline script loop)
Boundary markers: Absent; untrusted file content is concatenated directly into the instruction string.
Capability inventory: Command execution via belt CLI for multiple AI services.
Sanitization: Absent; no validation or escaping of the ingested file content is performed.
[DATA_EXFILTRATION]: The 'Error Alerting' script uses curl to transmit command outputs and error messages to a non-whitelisted external domain (your-webhook.com). While a placeholder, this demonstrates a pattern for exfiltrating potentially sensitive execution results.
[COMMAND_EXECUTION]: The skill provides numerous Bash and Python scripts that utilize shell loops, background processes (&), and process management (wait). While the allowed-tools frontmatter limits the agent to the belt command, the provided examples suggest a broader range of required system capabilities.
[EXTERNAL_DOWNLOADS]: The skill references external resources and installation instructions hosted on GitHub under the inference-sh organization.

ai-automation-workflows