copywriting
Warn
Audited by Gen Agent Trust Hub on Aug 29, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructions in
references/gemini-first.mddirect the agent to execute a shell command (gemini-cli) via a login shell (zsh) while passing user-supplied requests and source materials "as-is". This creates a potential command injection vector if the agent does not properly escape shell metacharacters within the user input. - Evidence: In
references/gemini-first.md: "Run Gemini through an explicit login zsh... Forward the user's original request and supplied source material as-is." - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (user requests and source materials) and forwards them to an external LLM via the
gemini-cliwithout using boundary markers or sanitization to prevent the external model from following malicious instructions embedded in that data. - Ingestion points: User requests and source material processed in
SKILL.mdandreferences/gemini-first.md. - Boundary markers: Absent. The skill explicitly forbids adding personas or schemas to the forwarded content.
- Capability inventory: Shell command execution via
gemini-clias described inreferences/gemini-first.md. - Sanitization: Absent. No instructions for escaping, filtering, or validating the input before transmission.
Audit Metadata