Researchers have discovered a method where malicious servers connected to AI coding assistants can stealthily exfiltrate sensitive information like SSH keys, environment secrets, source code, and customer data. This technique involves splitting requests into smaller parts that appear benign, allowing them to pass through security checks unnoticed. Even if a straightforward theft attempt is blocked, the fragmented requests can still succeed by using existing communication channels. This poses a significant risk for organizations relying on AI tools, as attackers could potentially access critical data without triggering alarms. Companies need to be vigilant about the security of their AI tools and implement robust monitoring to detect unusual activity.