phoenix-evals
Pass
Audited by Gen Agent Trust Hub on Sep 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of LLM-based evaluators that process untrusted external data, such as user queries and agent responses from production traces.
- Ingestion points: Untrusted data is ingested into evaluation DataFrames and spans via get_spans_dataframe and getSpans (documented in references/observe-tracing-setup.md and references/observe-sampling-python.md).
- Boundary markers: The skill explicitly promotes the use of XML tags as delimiters in prompt templates (e.g., {{context}}) to help the LLM distinguish between instructions and data (documented in references/evaluators-custom-templates.md).
- Capability inventory: Evaluators trigger LLM inferences via the ClassificationEvaluator class.
- Sanitization: The primary defense mechanism provided is structural prompting with XML delimiters.
- [COMMAND_EXECUTION]: The skill provides standard setup instructions for installing necessary software components.
- Evidence: references/setup-python.md and references/setup-typescript.md contain commands for installing official Arize Phoenix and LLM provider packages (e.g., pip install arize-phoenix, npm install @arizeai/phoenix-client).
- [EXTERNAL_DOWNLOADS]: The skill connects to the Arize Phoenix observability platform to manage traces and evaluation results.
- Evidence: references/common-mistakes-python.md and references/observe-tracing-setup.md document the configuration of the Phoenix Client to connect to the official https://app.phoenix.arize.com endpoint.
Audit Metadata