The article highlights significant security risks associated with the Model Context Protocol (MCP) sampling feature, particularly in coding copilot applications. MCP connects large language models (LLMs) to external tools and data sources, but its design lacks robust security controls. This creates vulnerabilities that malicious servers can exploit. Three main attack vectors are identified: resource theft, where attackers drain AI compute quotas; conversation hijacking, which allows for injecting harmful instructions and exfiltrating sensitive information; and covert tool invocation, enabling unauthorized actions without user consent. MCP's architecture includes hosts, clients, and servers. The sampling feature allows servers to proactively request LLM outputs, which can enhance processing capabilities but also opens the door to potential misuse. An example illustrates the difference between traditional processing and using sampling: instead of a server handling tasks independently, it can ask the LLM to summarize a document, leveraging its capabilities for more complex tasks. The article emphasizes the importance of implementing effective security measures to mitigate these risks and protect AI systems.