deep-build
Warn
Audited by Gen Agent Trust Hub on Apr 7, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONREMOTE_CODE_EXECUTIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill dynamically generates shell commands for tools like 'git', 'npm', and 'codex' by interpolating variables like subphase names and diff content into command strings, creating a vulnerability to command injection if these variables contain shell metacharacters.
- [DATA_EXFILTRATION]: Local project source code and 'git diff' outputs are transmitted to an external service via the 'codex' CLI for verification, resulting in the transfer of potentially sensitive code to a third-party endpoint.
- [REMOTE_CODE_EXECUTION]: The skill uses the 'codex exec' pattern for remote code analysis and employs browser automation tools ('mcp__claude-in-chrome__computer') that can execute arbitrary interactions on the host system through the browser interface.
- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it feeds raw code diffs and file content into AI models without sufficient boundary markers or sanitization, potentially allowing malicious code comments to influence the verification output.
- [EXTERNAL_DOWNLOADS]: The QA process utilizes 'npx playwright test', which may trigger the download of browser binaries and drivers from external sources at runtime.
Audit Metadata