improve-codebase-architecture
Pass
Audited by Gen Agent Trust Hub on Aug 26, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The skill ingests untrusted data from the local codebase, including source files, ADRs, and git logs. This creates a surface for indirect prompt injection where malicious instructions in the code could influence the agent. \n
- Ingestion points: Local codebase files,
CONTEXT.md, thedocs/adr/directory, and output fromgit log --oneline. \n - Boundary markers: The skill lacks explicit delimiters to distinguish untrusted codebase data from its own instructions. \n
- Capability inventory: File system writes to temp directories, shell command execution for opening reports, sub-agent spawning, and modification of project files like
CONTEXT.md. \n - Sanitization: The Mermaid diagramming configuration uses
securityLevel: \"loose\", which enables HTML rendering within diagrams and increases the risk of XSS if the agent includes unescaped content from the codebase in the report. \n - [COMMAND_EXECUTION]: The skill invokes platform-specific shell commands (
xdg-open,open, orstart) to automatically open the generated HTML review for the user. \n - [EXTERNAL_DOWNLOADS]: The generated HTML report references external assets from well-known services, including Tailwind CSS from
cdn.tailwindcss.comand Mermaid fromcdn.jsdelivr.net.
Audit Metadata