paul-graham-perspective
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill's 'Agentic Protocol' in
SKILL.mdcreates a surface for indirect prompt injection by instructing the agent to gather data from external search results. However, the skill provides a multi-step reasoning framework including 'Checkpoint' self-tests and internal summarization steps that mitigate the risk of search results directly influencing the agent's behavior. - [COMMAND_EXECUTION]: The
SKILL.mdfile includes instructions in the '版本自检 (静默)' section for the agent to execute localgitcommands (rev-parseandls-remote) to check for skill updates. These commands are limited to the skill's own directory and serve a benign maintenance purpose. - [EXTERNAL_DOWNLOADS]: The
README.mdfile recommends installing the skill usingnpx skills, which fetches content from the author's GitHub repository. This utilizes a standard deployment utility and represents legitimate software distribution within the platform's ecosystem. - [PROMPT_INJECTION]: The skill contains prescriptive role-playing instructions to ensure persona consistency and specifically warns against common 'hedging' behaviors. It follows security best practices for persona skills by providing a dedicated 'EXIT TRIGGER' in
SKILL.mdto allow users to return the agent to its default state.
Audit Metadata