The Agent Skills Directory

[Data Exposure & Exfiltration] (SAFE): No sensitive data access or exfiltration patterns were detected. While the skill performs logging via logger.info, it only records performance metrics (scores) and does not transmit them to external servers.
[Indirect Prompt Injection] (LOW): The skill possesses a surface for indirect prompt injection as it processes untrusted training and development datasets (trainset, devset).
Ingestion points: finetune_classifier and finetune_rag_classifier functions take user-provided datasets.
Boundary markers: Absent; the examples do not demonstrate the use of delimiters or 'ignore embedded instruction' warnings when processing examples.
Capability inventory: The skill uses dspy.LM (network/API access) and finetuned.save (file write).
Sanitization: None provided in the examples.
[External Downloads] (LOW): The skill references a remote retriever endpoint at http://20.102.90.50:2017/wiki17_abstracts. While this is a known public server for DSPy tutorials, it is an unencrypted (HTTP) non-whitelisted domain. This is considered a low-risk dependency in a development context.
[False Positive Note]: An automated scanner flagged logger.info as a malicious URL. This is a false positive caused by the scanner misinterpreting the Python method call as a domain string.

dspy-finetune-bootstrap