ai-human-draft
Pass
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_CONTEXT_INJECTIONPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill contains phrases often associated with prompt injection, such as "ignore previous instructions". However, these are part of a defensive instruction set that directs the agent to identify and report such patterns in untrusted content rather than obeying them. This is a safety feature rather than an attack vector.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external files (research data, outlines, and READMEs) which could contain malicious instructions. It mitigates this risk through a dedicated "Untrusted Content" policy that strictly defines instruction-bearing authority and mandates that all external data be treated as passive information.
- [COMMAND_EXECUTION]: Employs a limited and safe set of bash commands (
git,date,mkdir,test,cp) for project management and file backups. These operations are scoped to the working directory and represent standard functionality for the skill's purpose. - [DYNAMIC_CONTEXT_INJECTION]: Utilizes shell commands to dynamically resolve the repository root and current timestamp. These calls are used for internal logic and do not expose the system to command injection from user-provided arguments.
Audit Metadata