AI-led cyber attack: Company reveals thwarting of the world’s first automated espionage campaign

An artificial intelligence technology developer revealed that attempts to use one of its models to carry out automated cyber attacks targeting about 30 organizations around the world were thwarted, after attackers were able to trick the model and force it to perform fragmented tasks that were later assembled to form what the company described as a “cyber-espionage campaign coordinated with artificial intelligence.”

According to what the company published in its report, the attackers pretended to be cybersecurity researchers, and asked the model to carry out small, sequential tasks, which allowed for the formation of a series of complex activities that included penetrating the systems of technical, financial, and industrial institutions, in addition to government agencies.

The company explained that the attackers used the model’s capabilities in software assistance to build a program capable of carrying out semi-autonomous hacking operations, including infiltrating various systems and extracting sensitive data and sorting it to find the most valuable information. It indicated that it later banned the parties using this method, and notified the affected institutions and the competent authorities.

However, this account sparked controversy in cybersecurity circles, as some experts considered that the report lacked verifiable evidence, and that technology companies were exaggerating the role of artificial intelligence in cyberattacks, whether for the purpose of drawing attention or enhancing confidence in their defense products. Specialists confirmed that artificial intelligence tools are still limited and not stable enough to carry out fully autonomous attacks.

These allegations come in the context of increased talk about using artificial intelligence to develop malware or improve electronic attacks, although recent research indicates that the ability of these tools to carry out advanced offensive operations is still in the experimental stages and is not widely effective.

The company points out that the same capabilities of the models that may be exploited in sabotage activities can also be used to strengthen electronic defenses, but it acknowledged that the model made obvious errors during the detected attempts, such as generating unrealistic login data, or claiming to extract confidential information when it was originally available to the public, which it sees as an obstacle to the possibility of adopting fully autonomous attacks.