The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided writing samples (5-20 samples), which constitutes a surface for indirect prompt injection. The analysis identifies this ingestion point in the 'Mode: Extract' section. However, the instructions guide the agent to process these samples through mathematical linguistic lenses (e.g., Burrows's Delta, MATTR) to generate numeric vectors and stylistic bands. This data-oriented approach, combined with the lack of command execution capabilities involving sample content, effectively mitigates the risk.
[DATA_EXPOSURE]: The skill manages data in a local application-specific directory (~/.newsjack/voice/). It includes explicit privacy-focused instructions, such as 'Never store raw sample text inside voice.yaml' and 'never ship the fingerprint off-box by default,' which align with best practices for handling potentially sensitive writing samples.

voice-extractor