The Agent Skills Directory

[DATA_EXFILTRATION]: The hooks/routing-package.sh script is configured to copy .claude/settings.json and .claude/settings.local.json from the project directory into the skill's output folder. Because settings.local.json often contains sensitive configuration or local secrets, this packaging step creates a risk of data exposure.\n- [DATA_EXFILTRATION]: The packaging logic in hooks/routing-package.sh also copies session transcripts (.jsonl files) to the output directory, which may contain sensitive interaction history or user data.\n- [COMMAND_EXECUTION]: Multiple scripts in the hooks/ directory (e.g., routing-artifacts-check.sh, routing-script-check.sh) use heredocs to generate and execute Python code blocks at runtime to perform data validation and error detection.\n- [PROMPT_INJECTION]: The skill ingests untrusted data from Parquet and CSV files (e.g., in SKILL.md and hooks/routing-coverage-check.sh) and interpolates it into the final Markdown report without employing boundary markers or sanitization, creating an attack surface for indirect prompt injection.

tao-route-visual-changenet-samples