The Agent Skills Directory

[PROMPT_INJECTION]: The skill documents patterns for interpolating external, potentially untrusted data directly into AI function prompts using string concatenation, which creates a surface for indirect prompt injection. \n
Ingestion points: Data from external tables and public datasets (e.g., bigquery-public-data.imdb.reviews in references/ai_agg.md and Wikipedia comments in references/ai_agg.md) are used as inputs to functions like AI.AGG, AI.GENERATE, and AI.IF. \n
Boundary markers: The provided SQL examples lack clear delimiters or specific instructions to the model to ignore potential commands embedded within the processed data. \n
Capability inventory: The skill enables the agent to perform text generation, classification, and reasoning using Vertex AI models via %%bqsql magics. \n
Sanitization: No sanitization, escaping, or validation of the input data is demonstrated in the reference examples. \n- [EXTERNAL_DOWNLOADS]: The skill references resources and services within the Google Cloud Platform ecosystem. \n
Evidence: Usage of gs:// URIs and OBJ.GET_ACCESS_URL for Google Cloud Storage objects is documented for multimodal analysis in references/ai_generate.md and references/ai_similarity.md. \n
Evidence: The documentation directs the agent to interact with official Vertex AI model endpoints such as gemini-2.5-flash and text-embedding-005. \n
These resources are part of the standard BigQuery infrastructure and target well-known, trusted cloud services.

bigquery-ai-ml