The Agent Skills Directory

[PROMPT_INJECTION]: The ai_screening component is vulnerable to indirect prompt injection. The skill interpolates raw, untrusted data ({resume_text}) directly into an LLM prompt to determine candidate advancement or rejection.
Ingestion points: The {resume_text} variable in SKILL.md (AI Resume Screening section) processes external content from job applicants.
Boundary markers: None. The resume content is placed directly inside the prompt without delimiters or instructions to ignore embedded commands.
Capability inventory: The output of this screening directly triggers actions like advance_to_screening, auto_reject, and send_rejection.
Sanitization: There is no evidence of sanitization or filtering of the resume content before it is processed by the model.
[DATA_EXFILTRATION]: The skill is designed to process and collect highly sensitive information, which represents a significant data exposure surface.
Sensitive Data Handling: The workflow explicitly collects and manages tax_forms, direct_deposit (banking) information, and government id documents within the onboarding_workflow and document_compliance sections.
Network Exposure: Data is transmitted to multiple external platforms including Airtable, Slack, and various HRIS APIs (Greenhouse, Workday, BambooHR).

hr-automation