databricks-dbsql

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.65). The skill’s runtime workflow can ingest outsider-authored free text via read_files(...) (e.g., reading arbitrary JSON/text files from external storage paths) and then pass that content into LLM-invoking functions like ai_query()/ai_parse_document() for extraction/summarization, creating an indirect prompt-injection path from file contents into the agent’s LLM context.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The skill reads external files at runtime (e.g., READ_FILES('/Volumes/catalog/schema/volume/invoices/' ) and similar s3:// paths) and then injects the fetched document content into ai_query prompts (CONCAT('Extract ...', doc:document:elements[0]:content::STRING)), which means remote file content can directly control model prompts; this is a runtime external dependency that could enable prompt-injection risks.