improve-codebase-architecture
Warn
Audited by Gen Agent Trust Hub on Jun 22, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [DYNAMIC_EXECUTION]: The generated HTML report template in
HTML-REPORT.mdinitializes the Mermaid library withsecurityLevel: "loose". This setting enables the rendering of arbitrary HTML and the execution of JavaScript within diagrams. Because the skill incorporates content from the local codebase (file names, module descriptions, etc.) into these diagrams without explicit sanitization, a codebase containing malicious metadata could trigger local script execution when the report is opened in a browser. - [COMMAND_EXECUTION]: The skill uses shell-specific commands (
xdg-open,open, orstart) to automatically launch the user's default web browser and display the generated report from the temporary directory. - [EXTERNAL_DOWNLOADS]: The HTML report template fetches external resources at runtime from
cdn.tailwindcss.comandcdn.jsdelivr.net. While these are well-known technology services, they represent a dependency on external infrastructure for the report's functionality. - [DATA_EXFILTRATION]: The skill systematically reads sensitive architectural data, including domain models (
CONTEXT.md), architectural decision records (ADRs), and the codebase structure, writing this information to files in the system's temporary directory. - [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8). It ingests untrusted data from the codebase being analyzed (Ingestion point: codebase exploration tool) and interpolates this data into the HTML report and the 'Grilling' dialogue without documented boundary markers or sanitization procedures. This data is then used to drive agent decisions during the architectural review process.
Audit Metadata