The Dutch Institute for Vulnerability Disclosure (DIVD), an independent Dutch CERT focused on identifying and reporting weaknesses in digital systems, recently fell victim to a cyberattack that involved an autonomous AI agent. The attack began when the intruder exploited a security flaw to gain initial access. Once inside, the attacker reportedly handed over part of the operation to an AI tool, which then analyzed the outcomes of its actions and independently decided how to proceed within the network. This AI agent took control of certain aspects of the intrusion without waiting for human instructions, making it a more advanced and self-directed form of attack. The traces left behind by the AI suggest that the tool was capable of performing an action, evaluating the results, and quickly deciding on the next step based on what it had learned. It could reassess the situation, change tactics, or continue down a path that was working, all in real time. The DIVD describes this as an "agent-based" attack, where the AI wasn't just used to generate code or suggest commands, but actually took charge of part of the breach. What's unusual is that the AI also left behind detailed comments explaining its actions and the decisions it made during the attack, allowing researchers to trace its movement through the network. The DIVD described the operation as "very, very disorganized," happening at high speed but marked by questionable decisions. For example, the AI reportedly launched a password spraying campaign—where it attempts to log in with a few common passwords across many accounts—while simultaneously conducting a man-in-the-middle (MITM) attack, designed to intercept and manipulate authentication processes. This combination of tactics is said to have disrupted the MITM attack itself, creating unnecessary complications. The DIVD noted several similar errors during the intrusion, suggesting that the AI may have been poorly trained or misconfigured for this type of operation. Despite these mistakes, the attack was serious enough to warrant a formal investigation. The DIVD has informed the Dutch police, the data protection authority, and the National Cyber Security Centre (NCSC). However, the details of the exploited system and the specific vulnerability remain confidential for now, to prevent potential harm to other users who might be similarly affected.