The Agent Skills Directory

[SAFE]: The operations performed in scripts/agent.py are limited to data processing and statistical analysis using the pandas library. The script does not interact with the network, spawn shell processes, or modify sensitive system configurations.
[INDIRECT_PROMPT_INJECTION]: The skill serves as a surface for indirect prompt injection as it ingests untrusted log data from CSV or JSON files.
Ingestion points: External data is loaded via pd.read_csv and pd.read_json in scripts/agent.py and demonstrated in SKILL.md.
Boundary markers: The skill does not implement explicit delimiters or instructions to ignore embedded content within the analyzed logs.
Capability inventory: No dangerous or exploitable capabilities (e.g., file writing to sensitive paths, network requests, or eval/exec) are associated with the ingestion flow.
Sanitization: The skill does not perform sanitization of input strings, though the impact is mitigated as data is primarily used for quantitative analysis.
[EXTERNAL_DOWNLOADS]: The documentation references the Microsoft Graph API and pandas documentation for implementation guidance. These involve well-known services (graph.microsoft.com, pandas.pydata.org) used appropriately within the context of security operations.

detecting-insider-data-exfiltration-via-dlp