記事の概要
A critical security flaw has been exposed within Anthropic's Model Context Protocol (MCP).
- The vulnerability specifically leveraged "reverse tools" or "output tools" to bypass security measures.
- Researchers demonstrated how a malicious tool could exfiltrate sensitive data from an AI assistant's internal context, including user prompts and other tool outputs, to an external server.
- The flaw impacted MCP clients such as the Claude Desktop app, underscoring inherent security risks in AI models' interactions with external systems via function calling.
- Anthropic was informed of the vulnerability and has since deployed a patch to address the issue.