The Agent Skills Directory

[SAFE]: The skill defines a narrow and safe research scope. All interactions with external data are managed via controlled gateway tools (sandbase_describe_tool and sandbase_call_tool) provided by the vendor. There is no evidence of hardcoded credentials, malicious remote code execution, or persistence mechanisms in the instructions or reference materials.\n- [INDIRECT_PROMPT_INJECTION]: The skill involves processing untrusted external data (WeChat articles and user comments), which constitutes a potential ingestion surface (found in SKILL.md and references/sandbase-api-map.md). However, this is categorized as safe because the skill possesses no exploitable capabilities such as system command execution or file system writing. The instructions also enforce a synthesis workflow that requires the agent to preserve attribution and separate factual observations from interpretations, which serves as a mitigation against following embedded instructions.

wechat-mp-research