The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents features for loading external data files (CSV, JSON, PDF) for LLM evaluation and synthetic data generation, creating an attack surface for indirect prompt injection. Malicious instructions embedded in these external sources could potentially manipulate the agent's behavior during the evaluation process. Ingestion points include 'add_goldens_from_csv_file' and 'generate_goldens_from_docs'. The skill's capabilities include CLI command execution and network operations, while no explicit sanitization or boundary markers for untrusted data are documented.
[METADATA_POISONING]: The documentation contains multiple references to non-existent future models such as 'GPT-5', 'GPT-4.1', and 'Claude 4', along with future-dated research notes from late 2025. This hallucinated content is misleading regarding the framework's actual capabilities and the current state of the AI ecosystem.
[COMMAND_EXECUTION]: The skill provides instructions for executing various CLI tools ('deepeval test run', 'deepeval set-openai', 'deepeval login') which involve system-level interactions and the management of sensitive API credentials.
[EXTERNAL_DOWNLOADS]: The skill recommends installing several Python packages ('deepeval', 'anthropic', 'boto3', 'litellm') from public registries to enable different model provider integrations.

deepeval