test-suite-architect
Pass
Audited by Gen Agent Trust Hub on Nov 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The core functionality of the skill involves an autonomous execution loop where the AI agent is directed to read test case specifications from external markdown files (e.g.,
tests/docs/02-CLI-TEST-CASES.md) and "Execute all steps exactly as documented" (evidence inreferences/master_qa_prompt.md). This architecture is inherently vulnerable to indirect prompt injection, where malicious instructions placed in a project's test files could be executed by the agent.\n - Ingestion points:
tests/docs/*.md(test case documents).\n - Boundary markers: Absent. The instructions explicitly command the agent to not deviate from the external documentation.\n
- Capability inventory: The agent is granted shell access with capabilities for repository cloning (
git), dependency management (pnpm), container interaction (docker), and skill installation (ccpm).\n - Sanitization: No sanitization or validation of the content within the test case documents is performed before execution.\n- [COMMAND_EXECUTION]: The skill's documentation and prompts (
references/day1_onboarding.md,references/llm_prompts_library.md) provide a wide array of shell commands for the agent to perform. These include environment setup commands such asgit clone,pnpm install, anddocker exec, as well as testing commands likeccpm installandcurl. While intended for legitimate QA activities, they represent a powerful set of capabilities that could be abused if the agent is successfully injected.\n- [PROMPT_INJECTION]: Thereferences/master_qa_prompt.mdandreferences/llm_prompts_library.mdfiles contain role-play instructions ("You are a senior QA engineer with 20+ years of experience at Google") and authoritative directives ("CRITICAL INSTRUCTIONS", "MANDATORY RULES", "DO NOT skip tests"). These are designed to override the agent's default behavior and ensure strict adherence to the skill's testing workflow, mimicking patterns often found in prompt injection attacks.
Audit Metadata