smart-search
Pass
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute the 'opencli' command-line tool with user-influenced arguments. It encourages the construction of queries using user-supplied parameters such as '主题' (Topic) and '目标' (Goal). If the agent platform does not properly escape these inputs before shell execution, it creates a potential surface for command injection.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to fetch and summarize data from numerous untrusted external platforms, including social media (Twitter/X, Weibo, Reddit), forums (Zhihu, V2EX), and news sites (Reuters, 36kr).
- Ingestion points: External data enters the agent context through the output of 'opencli' commands targeting the domains specified in the supporting reference files.
- Boundary markers: The instructions do not define any delimiters or provide safety warnings to the agent to ignore instructions that might be embedded within search results.
- Capability inventory: The agent has the capability to execute shell commands (via 'opencli') and provide comprehensive responses to users based on the retrieved information.
- Sanitization: There are no instructions for filtering, escaping, or validating the content retrieved from external websites before it is processed by the AI.
Audit Metadata