Hackers Are Now Using AI to Run Attacks

Researchers found the first real-world attack where an AI automates tasks after a breach, a major shift in cyber threats.

Security researchers at Kaspersky have identified what they believe is the first documented use of a large language model (LLM) to automate part of a live cyberattack. According to a report from its SecureList blog, a campaign dubbed CyberStrikeAI targeted organizations in the education sector across Africa and Latin America. While the initial method of entry was not detailed, the key discovery was how the attackers behaved once inside a network. Instead of a human operator manually issuing commands, the attackers deployed an AI-powered agent to control their malware. This agent was responsible for the post-exploitation phase, which includes everything an attacker does after gaining initial access, such as exploring the network, identifying valuable data, and establishing persistent access. This finding represents a significant and long-feared evolution in the capabilities of threat actors, moving AI from a preparatory tool to an active participant in an intrusion.

Technically, the innovation lies in automating the command-and-control (C2) process. After a system is compromised, malware typically “calls home” to a C2 server to receive instructions from the attacker. In a traditional attack, a human operator analyzes the compromised environment and decides what to do next, typing commands to steal files or move to other machines. The CyberStrikeAI campaign replaced this human operator with an AI model. This AI could likely interpret data from the victim’s system and dynamically generate its own commands to explore the network or escalate privileges. This is far more advanced than a simple automated script, which can only follow a predefined set of instructions. An LLM-powered agent can adapt to unexpected situations, making it more evasive and efficient than both human attackers and rigid scripts.

Until now, the conversation around the malicious use of AI in cyberattacks has centered on its role in preparation and social engineering. Security experts have demonstrated how LLMs can be used to generate highly convincing phishing emails, write polymorphic malware that changes its code to avoid detection, or discover vulnerabilities in software. These are all serious threats, but they still require a human to execute the main phase of the attack. This discovery confirms that threat actors have crossed a new threshold. The use of AI is no longer limited to crafting the lure; it is now being used to pilot the attack itself. This shift from a strategic assistant to a tactical operator was a theoretical risk that security professionals have been warning about for years, and its arrival in the wild marks a pivotal moment for the cybersecurity landscape.

For founders, CTOs, and security teams, this development demands a fundamental shift in defensive thinking. Attacks driven by AI can operate at machine speed, 24 hours a day, without rest. An AI agent can test thousands of potential attack paths in the time it would take a human to evaluate a few, drastically shortening the “dwell time” between initial breach and mission completion. This acceleration means that human-led incident response may be too slow to stop a major data breach. The new reality is that organizations will increasingly need AI-powered defensive tools to fight back. Automated threat detection and response systems that can identify and neutralize anomalous activity in real-time will become essential. As this technology becomes more accessible, we can expect to see more threat groups adopting it, lowering the barrier for sophisticated, adaptive attacks and creating a new, faster-paced arms race between attackers and defenders.

This is the first confirmed use of an LLM to automate the post-exploitation phase of a live attack. It proves that AI can now actively manage command-and-control operations, moving beyond prep work and fundamentally changing the speed at which attackers can operate within a compromised network.

AI-automated attacks dramatically reduce the time between an initial breach and significant damage, like data theft. This increases business risk by giving security teams far less time to respond. Traditional, human-led security monitoring may no longer be sufficient to prevent major incidents.