The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The documentation file references/chunking-strategies.md includes a code snippet for LateChunker that uses AutoModel.from_pretrained(model_name, trust_remote_code=True). Enabling trust_remote_code allows the library to download and execute arbitrary Python code contained within the model's repository.
[EXTERNAL_DOWNLOADS]: The skill instructs the agent to download pre-trained machine learning models from external repositories like Hugging Face and to interface with various third-party APIs (OpenAI, Cohere, Voyage AI, etc.).
[DATA_EXFILTRATION]: The workflow involves reading local documents and sending their content to external cloud-based APIs for processing, which represents a potential path for data exposure.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection. Ingestion points: External documents are processed in SKILL.md and references/chunking-strategies.md. Boundary markers: None identified; document content is concatenated directly into prompts via string joining. Capability inventory: The skill uses network APIs (OpenAI, Cohere), vector database persistence, and file processing. Sanitization: No evidence of content filtering or safety instructions to ignore embedded commands.
[SAFE]: The skill follows security best practices by using placeholders for API keys and suggesting the use of environment variables for secret management.

rag-architect