bigquery-agent-analytics-sdk

Pass

Audited by Gen Agent Trust Hub on Jun 25, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill requires the installation of the bigquery-agent-analytics Python package and its optional dependencies (llm, bigframes) from external registries.
  • [COMMAND_EXECUTION]: The skill includes a CLI suite (bqaa or bq-agent-sdk) that executes commands for infrastructure management, data seeding, and evaluation pipelines.
  • [PROMPT_INJECTION]: The skill features 'LLM-as-Judge' and trajectory matching evaluators that process agent telemetry. This introduces an indirect prompt injection surface where untrusted data within agent traces could attempt to manipulate the judging model's output.
  • Ingestion points: Data enters the system through client.get_trace() and client.get_traces_for_session(), which are then passed to LLMAsJudge.evaluate_traces().
  • Boundary markers: No specific delimiters or safety instructions for the judge LLM are documented in the skill instructions.
  • Capability inventory: The SDK performs BigQuery queries, writes results to tables, and can deploy Google Cloud resources via the CLI.
  • Sanitization: There is no mention of sanitizing or escaping the trace content before it is processed by the evaluation LLM.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 25, 2026, 01:15 AM
Security Audit — agent-trust-hub — bigquery-agent-analytics-sdk