The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/native_process.py utilizes ctypes.CDLL(None) to interface with Linux-specific system calls (pidfd_open and pidfd_send_signal) for low-level process management during model inference. This is used to ensure that worker processes within containers are correctly terminated.
[PRIVILEGE_ESCALATION]: The utility scripts/restore_docker_mount_permissions.py uses docker run with volume mounts to perform recursive chown and chmod operations on host directories. This is designed to fix ownership issues for files created by root-level Docker processes but represents a capability to modify host access controls.
[COMMAND_EXECUTION]: Multiple bundled scripts, including scripts/prepare_cosmos_reason_model.py and scripts/batch_cosmos_embed_text.py, use the subprocess module to execute Docker containers and native CLI tools as part of the machine learning pipeline.
[EXTERNAL_DOWNLOADS]: The workflow fetches model configuration metadata from the official Hugging Face API and downloads model checkpoints from Hugging Face repositories using the huggingface_hub library.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection as it processes external datasets and model-generated predictions to drive training and evaluation loops.
Ingestion points: scripts/prepare_gap_analysis_predictions.py reads LLaVA annotations from kpi_dataset.annotations_path; scripts/compute_bcq_accuracy_metrics.py reads model results from results.json files.
Boundary markers: No explicit markers or "ignore instructions" warnings are used to delimit question text or model responses during ingestion or interpolation.
Capability inventory: The skill possesses capabilities for subprocess execution, Docker container management, and broad filesystem write access.
Sanitization: The skill lacks specific sanitization or filtering for potentially malicious instructions embedded within the question text or model outputs.

tao-run-deft-cr-its-mining