Stop asking 'is this skill safe?' — ask 'what can it do?'

The author highlights a critical security issue regarding the use of AI agents. The article describes a real-world scenario where a user accidentally restored a Trojan via a backup hidden within an AI agent's configuration file, disguised as a standard 'skill.' The core problem is that users often trust extensions and scripts without inspecting their content, even though the agent executes this code with user privileges. The author urges a shift in security assessment: instead of blind trust, one must analyze the functionality of every module. As a solution, an open-source tool is introduced that automatically analyzes skill code, allowing users to understand exactly what actions the agent is performing. This approach emphasizes the importance of transparency and control in an era of widespread autonomous AI systems, where malicious code can easily be masked as legitimate functionality.