The Agent Skills Directory

Indirect Prompt Injection Surface: The AI.GENERATE function documented in references/ai_generate.md processes content from database tables by incorporating it directly into model prompts. This is a common pattern for processing structured and unstructured data with generative models.
Ingestion points: Data from table columns, such as article_content and invoice_text, are used as inputs for generation in references/ai_generate.md.
Boundary markers: Examples show string concatenation; implementing delimiters or specific instructions to ignore embedded content is a standard practice for production workloads.
Capability inventory: The function facilitates content generation, summarization, and structured data extraction through BigQuery's native Vertex AI integration.
Sanitization: The examples focus on SQL syntax; data-level sanitization would typically be handled at the application or data pipeline layer.
External Resource Integration: The documentation demonstrates using BigQuery object tables and connections to access Google Cloud Storage buckets (e.g., gs://cloud-samples-data/). These operations use established cloud identity and access management (IAM) permissions within the BigQuery environment.

bigquery-ai-ml