The Agent Skills Directory

[DATA_EXFILTRATION]: The skill is designed to process sensitive local transcripts. It implements several safeguards to prevent exfiltration:
File names are replaced with anonymized identifiers (e.g., 'own-01') in all reports.
Exact file lengths are bucketed into ranges (e.g., '0-5k') to prevent data linkage.
The '_force_local' function explicitly overrides configurations to prevent the use of cloud-based LLM providers, pinning execution to local tools like Ollama.
[PROMPT_INJECTION]: The tool ingests untrusted text from user-provided files which are then processed by a local LLM for PII detection. This creates a surface for indirect prompt injection.
Ingestion points: 'scripts/audit.py' reads files with '.md' or '.txt' extensions from a target directory.
Boundary markers: The script relies on the external 'confide_core' module for prompt construction and boundary management.
Capability inventory: Execution is limited to reading files, writing aggregate statistics to disk, and calling local NLP/LLM services.
Sanitization: The final report contains only counts and percentages; detected PII values and original text segments are explicitly omitted.

audit