tao-finetune-cosmos-reason
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses Python helper scripts to wrap CLI tools such as Docker, srun, and ffprobe. Analysis shows that these commands are constructed using argument lists passed directly to subprocess.run or shlex.quote for shell strings, which prevents command injection from user-supplied paths.
- [DYNAMIC_EXECUTION]: Orchestration modules generate Python pre-flight code and shell scripts to perform environment attestation (e.g., verifying GPU driver compatibility) inside containers. These executions are author-controlled and do not ingest untrusted logic.
- [EXTERNAL_DOWNLOADS]: The skill fetches model checkpoints from Hugging Face Hub. It employs verified libraries (huggingface_hub) and enforces immutable commit SHAs to ensure data integrity. No downloads from untrusted third-party servers were found.
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied video datasets and annotation files, which is a standard ingestion surface for fine-tuning tools. The associated risks are handled by inherent LLM safety protocols and the skill's primary context as a supervised training workflow.
Audit Metadata