deepeval-evaluation
Pass
Audited by Gen Agent Trust Hub on Nov 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill provides patterns for evaluating LLM outputs using metrics such as GEval and FaithfulnessMetric. This creates an attack surface for indirect prompt injection, where malicious instructions embedded in evaluated content (actual_output) could influence the judge model's evaluation logic. \n
- Ingestion points: Untrusted data is ingested from 'data/nightly-capture.json' and 'data/goldens.jsonl', as referenced in the test suites within the evaluation tasks. \n
- Boundary markers: The instructions do not explicitly mandate the use of boundary markers or instructions to ignore embedded content when passing data to the LLM judge. \n
- Capability inventory: The skill utilizes 'pytest' and 'deepeval' to execute tests and communicate with remote LLM providers for scoring. \n
- Sanitization: No evidence of sanitization or filtering of the evaluated text was found in the skill's instructions. \n- [COMMAND_EXECUTION]: The skill instructions involve shell commands such as 'pip install', 'deepeval login', and 'deepeval test run'. While standard for this toolset, they involve the execution of external binaries and network interactions. \n- [EXTERNAL_DOWNLOADS]: The skill documentation recommends installing 'deepeval' and 'pytest' from public package registries. These are established, well-known packages required for the skill's intended purpose.
Audit Metadata