tao-finetune-clip
Pass
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill's primary function involves processing user-supplied image-caption datasets and prompt files for fine-tuning and inference tasks.
- Ingestion points: Image directories, caption text files, and prompt files provided in the specification overrides (SKILL.md).
- Boundary markers: There are no documented delimiters or specific instructions to the agent to ignore embedded commands within the ingested text files.
- Capability inventory: The skill executes various CLI tools such as
clip trainandclip inferencethat interact with the file system and results directories, typically running within Docker containers (references/skill_info.yaml). - Sanitization: No explicit content filtering or sanitization steps are documented for the input text data.
- [DYNAMIC_EXECUTION]: The skill instructs the user to enable a specific environment variable (
TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1) to facilitate loading metadata from Lightning checkpoints, which involves Python unpickling. - The instructions include an explicit warning against using this setting for untrusted checkpoints, mitigating the risk of arbitrary code execution from malicious model files (SKILL.md, references/error-patterns.md).
- This functionality is a core requirement for handling checkpoints within the TAO ecosystem.
Audit Metadata