chinese-app-benchmark
Warn
Audited by Gen Agent Trust Hub on Apr 12, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructions involve executing several shell commands, including 'grep', 'pnpm build', and 'git push --no-verify'. The use of '--no-verify' is a concern as it explicitly bypasses client-side git hooks that typically enforce security, linting, or testing policies.
- [DATA_EXFILTRATION]: The skill is instructed to read the entire repository structure (package.json, router files, components) and query the live database to analyze activity. When combined with the 'git push' capability, this creates a potential path for sensitive repository information or database data to be exfiltrated through modified commits.
- [PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection. In Phase 2B, it performs web searches for competitor UX analysis. Maliciously crafted content on external websites could provide instructions that influence the agent's implementation phase (Phase 4), potentially leading to the injection of malicious code into the repository.
- Ingestion points: Web search results and image search data (Phase 2B).
- Boundary markers: None present to distinguish between research data and instructions.
- Capability inventory: 'grep', 'pnpm build', 'git push', and live database querying.
- Sanitization: No validation or sanitization of external search content before it informs code implementation.
Audit Metadata