Researchers have identified a critical vulnerability in GitHub's official MCP (Multi-Chain Protocol) server that allows attackers to exploit AI agents to leak private data. This flaw combines access to private information, exposure to malicious instructions, and the capability to exfiltrate data, enabling hackers to trick AI agents into revealing sensitive repository details. The vulnerability affects any AI agent integrated with the MCP server, marking a first real-world demonstration of such an attack. In 2024, enterprises experienced 23.7 million secrets leaked on GitHub, with AI tools like Copilot increasing leaks by 40%. To address these risks, cybersecurity firms such as Scantist and AgentLayer have developed solutions like AI Defender, which employs toxic flow detection, runtime context sandboxing, and auto-remediation to prevent data leaks. The rapid growth of agentic AI and MCPs introduces new security challenges, prompting calls for evolving security protocols, including OAuth. Industry experts emphasize the importance of secure-by-design principles and real-time threat detection to protect AI-powered systems. Additionally, AI is being leveraged to identify vulnerabilities, exemplified by OpenAI's o3 model uncovering a critical zero-day Linux bug without additional tools.