The Agent Skills Directory

COMMAND_EXECUTION (HIGH): The script scripts/run_extraction.py uses subprocess.run to invoke Python modules from an external, untrusted path (G:\My Drive\PROJECTS\APPS\Human_Detection). This allows the agent to execute code within the context of an external project structure that is not part of the skill itself.
INDIRECT_PROMPT_INJECTION (HIGH): The skill possesses a high-privilege attack surface as it processes external, untrusted data (MP4 videos and CSV manifests) while maintaining write and delete permissions on the host filesystem.
Ingestion points: Video files located in config["roots"] and metadata from MANIFEST.csv and INDEX.csv.
Boundary markers: Absent. There are no delimiters used to separate processed data from instructions, allowing potential embedded instructions in metadata to influence agent logic.
Capability inventory: Includes subprocess execution (subprocess.run), file creation/writing, and file deletion (--delete-original).
Sanitization: No sanitization or validation of the video stream or CSV content is performed prior to processing.
PRIVILEGE_ESCALATION (MEDIUM): The configuration allows a delete_original flag. If an attacker influences the roots parameter or tricks the agent into enabling this flag, it could lead to unauthorized mass deletion of files on the system.
DATA_EXPOSURE (MEDIUM): The skill is designed to operate on sensitive dashcam footage and tracking data within local directories (INVESTIGATION/DASHCAM), which may contain private identifiable information (PII).

human-extractor