The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates the execution of powerful administrative shell commands.\n
Evidence: drush user:create, drush user:login, drush config:set, and drush state:set allow for significant modification of the Drupal environment and user access control.\n- [DYNAMIC_EXECUTION]: Documents the use of drush php:eval, which executes arbitrary PHP code directly in the application context.\n
Evidence: Examples in SKILL.md show drush php:eval "code" and its use for debugging class existence. This allows for the execution of logic not predefined in the codebase.\n- [INDIRECT_PROMPT_INJECTION]: The skill features a surface for indirect prompt injection via non-interactive code generators.\n
Ingestion points: drush generate commands take --answers as JSON input (e.g., in SKILL.md), which an agent may populate from external or untrusted sources.\n
Boundary markers: There are no instructions to use delimiters or warnings to ignore instructions embedded in the provided data.\n
Capability inventory: The skill has access to file system writes (via generators), database access, and raw PHP execution.\n
Sanitization: No sanitization or validation logic is specified for the input data passed to these generators.\n- [DATA_EXFILTRATION]: Provides a command to dump the entire database, which contains highly sensitive information.\n
Evidence: drush sql:dump > dump.sql in SKILL.md allows for the extraction of all site data to a local file.

drupal-drush