clip-finetune-recipes
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted image-text data from external datasets, creating a surface for malicious instructions to potentially enter the agent context. 1. Ingestion points: Dataset loading via
create_webdataset_loaderandShardBuildermentioned inSKILL.md. 2. Boundary markers: Absent; no delimiters or instructions to ignore embedded content are used when handling dataset captions. 3. Capability inventory: Shell command execution viapython,torchrun, andpippackage installation. 4. Sanitization: No explicit sanitization or validation of captions is performed during the training pipeline. - [EXTERNAL_DOWNLOADS]: The skill facilitates the installation of the
clip-recipespackage from the public PyPI registry and supports fetching dataset shards from AWS S3 buckets.
Audit Metadata