coretax-automation-specialist
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from user-provided Excel and CSV files to automate the filling of tax forms. This creates a vulnerability surface where untrusted data could attempt to manipulate the agent's instructions or the JavaScript commands executed in the browser.
- Ingestion points: Processes data from files like
audit_logs/Audit_Trail_YYYYMM.xlsxand user-supplied Excel spreadsheets. - Boundary markers: The instructions lack specific delimiters or warnings to the agent to ignore instructions potentially embedded within the data values.
- Capability inventory: The agent has access to
chrome-devtools.evaluate_javascript,chrome-devtools.take_screenshot, anddesktop-rpa-computer-use(OS-level RPA). - Sanitization: No methods for sanitizing or escaping the values extracted from data files before they are interpolated into JavaScript strings are described.
- [DYNAMIC_EXECUTION]: The skill uses
chrome-devtools.evaluate_javascriptto dynamically interact with the web portal's DOM. This involves generating and executing code at runtime based on external input, which is a standard but high-capability automation pattern. - [COMMAND_EXECUTION]: The skill leverages browser-level and potentially system-level automation tools (
chrome-devtools-mcp,desktop-rpa-computer-use) to perform automated data entry and navigation within the host environment.
Audit Metadata