The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (income figures, marketplace forms, names, and SSNs) and instructs the agent to use browser automation tools (Playwright, Puppeteer) for filing. This creates an attack surface where malicious instructions embedded in processed tax documents could attempt to influence the agent's behavior during a filing session.
Ingestion points: Taxpayer identity data and financial figures are ingested in SKILL.md and detailed in the examples/ and references/ directories.
Boundary markers: The skill uses structured line-by-line questions and transcription instructions which provide some delimitation, though no formal sanitization logic is provided for text inputs.
Capability inventory: The filing.md file provides instructions for using browser automation tools (Playwright, Puppeteer, Selenium) to interact with the IRS Free File Fillable Forms website.
Sanitization: While the skill lacks formal input sanitization, it includes strong security guardrails in filing.md instructing the agent to never store sensitive PII (SSN, DOB, PIN) in logs or transcripts and to always require explicit user consent before submission.

schedule-2