multi-agent-client-onboarding
Warn
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: MEDIUMDATA_EXFILTRATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill explicitly instructs specialist agents to search for and inventory sensitive information such as "Connection strings", "API keys in configs", "OAuth configs", and "JWT usage" within client documentation and codebases. While the primary purpose is auditing, this functionality creates a significant risk of exposing credentials and internal secrets if the agents' findings are not strictly handled. A guardrail is mentioned specifically to prevent including a "Supabase token" in the report, confirming the system anticipates exposure of environment secrets.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from user-provided file paths and URLs that are then analyzed by three specialist agents. There are no explicit instructions for boundary markers or sanitization to prevent malicious instructions embedded within the analyzed documents (e.g., hidden in code comments or documentation) from influencing the agents' behavior during the audit.
- [COMMAND_EXECUTION]: Specialist agents are granted access to the
Bashtool and are directed to examine automation scripts, Makefiles, and CI/CD configurations. This capability, combined with the instruction to audit executable files, poses a risk that an agent could accidentally or intentionally execute untrusted code found in the analyzed repository.
Audit Metadata