The Agent Skills Directory

[DATA_EXPOSURE]: The skill provides explicit instructions to prevent the storage or logging of sensitive user data. Specifically, it mandates that the agent must never store EINs, EFTPS PINs, e-file PINs, or bank account information in logs or transcripts, instructing that these credentials should be used at filing time and then discarded. This aligns with security best practices for handling PII and financial credentials.
[REMOTE_CODE_EXECUTION]: While the skill mentions browser automation tools such as Playwright, Puppeteer, and Selenium, it does so in the context of describing how an agent can perform e-filing through authorized third-party providers. There are no patterns of downloading and executing untrusted scripts from the internet.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface as it ingests user-provided financial data and entity information to populate tax forms.
Ingestion points: User-provided entity name, EIN, address, and quarterly sales/liability data.
Boundary markers: The skill includes instructions to ask the user for explicit authorization and confirmation before any filing or payment actions occur.
Capability inventory: The skill is capable of performing mathematical computations for tax liability and guiding browser-based e-filing through third-party services.
Sanitization: The instructions include mandatory manual validation checks (math and sanity) to be performed before the deliverable is finalized.
[COMMAND_EXECUTION]: The skill describes procedural workflows for using browser automation for tax filing but does not contain arbitrary shell command execution or unsafe subprocess calls.
[EXTERNAL_DOWNLOADS]: The skill references official IRS resources (e.g., irs.gov/pub/irs-pdf/f720.pdf) and reputable tax filing services (e.g., Tax720.com, TaxBandits). These are well-known, trusted entities in the tax domain and do not represent a security risk.

form-720