tao-finetune-cosmos-reason
Pass
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches pre-trained model weights and architecture specifications from official Hugging Face repositories and retrieves container images from the NVIDIA NGC registry (nvcr.io).
- [INDIRECT_PROMPT_INJECTION]: The skill ingests user-supplied video datasets and annotations, which serves as a surface for indirect prompt injection.
- Ingestion points: Dataset annotations are loaded in
scripts/cosmos_common.py(load_annotation) and processed inscripts/train_video_conversation_single_gpu.py(VideoConversationDataset). - Boundary markers: The workflow establishes explicit validation gates and mandatory intake checks for record counts, overlaps, and structural family as documented in
SKILL.md. - Capability inventory: The skill utilizes
subprocess.runto orchestrate Slurm tasks, Docker containers, and media inspection utilities. - Sanitization: Employs
shlex.quotefor shell command assembly and uses structured configuration schemas (JSON/TOML) for execution parameters. - [DYNAMIC_EXECUTION]: Orchestration utilities dynamically generate Python and Shell scripts to manage containerized environment setup and model downloads. For example,
scripts/prepare_cosmos3_vlm_checkpoint.pyassembles container entrypoint scripts to handle model snapshots, andscripts/evaluation_workflow.pygenerates in-container probe scripts for runtime integrity attestation. - [COMMAND_EXECUTION]: Multiple helper scripts invoke
subprocess.runto interact with platform-level utilities including Docker, Slurm scheduler commands (sbatch/srun), and media processing tools likeffprobeorunsquashfs.
Audit Metadata