rag-architect
Warn
Audited by Gen Agent Trust Hub on Sep 22, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The file 'references/chunking-strategies.md' includes a 'LateChunker' implementation that utilizes 'AutoModel.from_pretrained(model_name, trust_remote_code=True)'. This flag allows the model to execute custom code defined in its repository upon loading, which is a potential vector for remote code execution if the source is not verified.\n- [EXTERNAL_DOWNLOADS]: Several code snippets demonstrate downloading external resources at runtime, including NLTK datasets ('nltk.download("punkt")') and machine learning models from Hugging Face through the 'transformers' and 'sentence_transformers' libraries.\n- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the design of RAG systems that ingest external, untrusted content, which is a known attack surface for indirect prompt injection.\n
- Ingestion points: Document data processed by 'RecursiveCharacterTextSplitter' in 'SKILL.md' and various chunking scripts in the references directory.\n
- Boundary markers: The implementation examples do not provide specific delimiters or ignore-instruction warnings for processed content.\n
- Capability inventory: The skill describes workflows that include network API requests, file system interactions via vector databases, and complex data processing.\n
- Sanitization: The provided templates lack explicit steps for sanitizing or filtering instructions embedded within document text.
Audit Metadata