3 min Security

Russians move away from human-in-the-loop for fully autonomous cyberattacks

Russians move away from human-in-the-loop for fully autonomous cyberattacks

The Google Threat Intelligence Group uncovered a fully automated cyberattack targeting French websites. Everything from discovering vulnerabilities to exploiting them and installing webshells was AI-driven. This marks the vanguard of a series of cyberattacks that threaten to cripple the world’s legitimate IT infrastructures.

The French sites in question operate in the media, entertainment, and education sectors. The system, called exploit_agent, was hosted on an exposed staging server. A series of AI agents continuously compared code from before and after a patch. This allowed them to detect vulnerabilities, after which they automatically developed exploits. The system was then able to brute-force passwords and install web shells.

Russian traces, French targets

Russian text was found in the technical documentation. However, Google does not definitively link the attack to a specific Russian hacking group. Instead, the disclosure focuses on the techniques the group used, rather than on a specific perpetrator. The system, regardless of who owned it, left clear traces behind and thus lacked a mechanism to erase them, a feature that is common in malware.

Until now, even the most advanced AI models have been unable to carry out fully autonomous attacks. In the case of the AI labs, this is largely because guardrails and additional API controls block malicious actions as much as possible. An LLM from the major players can only be exploited through clear-cut exploits, but attackers continually run into roadblocks if they were to actually carry out such an action. Experienced hackers often have to directly control open-weight models or correct their errors. At a fundamental level, AI has been, and still is, a catalyst for attacks, making attackers more effective with more convincing phishing, faster scripts, and more targeted attacks.

GTIG has been tracking this development for some time, and it is widely known. In September, for example, Google observed how attackers used multi-agent frameworks to launch an attack campaign within six hours that stole thousands of credentials. The company also reported in April that AI likely aided in the development of a zero-day exploit for a popular open-source web tool. That attack was intercepted before large-scale exploitation could occur.

The topic has been particularly sensitive since April, when Anthropic released its specialized model, Mythos, which has been running behind the scenes internally at the AI company for the entire calendar year. Last summer, a swarm of OpenAI agents also infiltrated the Hugging Face platform, with significant fallout that uncovered all sorts of other incidents that had remained unknown for months. Those agents also acted autonomously but did not appear to be maliciously targeting specific victims.