Description

AutoJack is a sophisticated exploit chain discovered in Microsoft AutoGen Studio, an open-source platform designed for developing and testing AI-powered multi-agent systems. The attack demonstrates how a seemingly harmless web page can manipulate an AI browsing agent and ultimately execute malicious commands on the host machine without requiring direct user interaction beyond providing a URL. The exploit combines three separate security weaknesses into a single attack path. First, the system's WebSocket implementation relied on origin validation that trusted localhost connections. While this restriction prevented ordinary external websites from connecting directly, a browser controlled by an AI agent could inherit localhost privileges and bypass the intended protection. Second, critical communication endpoints lacked proper authentication checks, allowing unauthorized access to sensitive functionality. Third, user-supplied parameters were insufficiently validated before being passed to command execution routines, creating an opportunity for operating system command injection. In a typical attack scenario, a developer runs AutoGen Studio locally and uses an AI agent capable of browsing web content. An attacker hosts a malicious webpage and convinces the user to submit its URL to the agent. Once the AI agent loads the page, embedded JavaScript establishes a connection to a local service, bypasses authentication requirements, and delivers a crafted payload. The application then processes the payload and launches attacker-controlled commands with the same privileges as the developer. Researchers demonstrated this behavior by remotely triggering a local application on the victim system. To address the issue, Microsoft strengthened parameter handling, removed insecure URL-based parameter passing, and ensured that all relevant endpoints are protected by standard authentication mechanisms. The vulnerability highlights a growing security challenge in AI agent ecosystems, where systems that interact with untrusted web content can inadvertently gain access to trusted local resources. Effective mitigation requires strong authentication, strict command allowlisting, secure isolation practices, and careful validation of all agent-controlled inputs.