SecurityNews

Researchers reconstruct AI-driven hacking campaign step by step

Researchers at Zscaler ThreatLabz have reconstructed how a single attacker deployed multiple AI models in an international hacking campaign. Each model was given its own task, from scouting targets to preparing stolen data for sale. The analysis comes shortly after the European Commission was hit by a cyberattack in which AI tooling played a role.

According to the Dutch tech news site Tweakers, the attackers behind the attack on the European Commission used the AI tool ARTEX to identify vulnerabilities. Several systems, including HR applications, were temporarily taken offline. A data breach involving personal data has not been ruled out.

The campaign analysed by ThreatLabz was discovered by its researchers themselves. It involved a customised implementation of ARTEX, which the attacker combined with Claude Code, Codex and Grok. Because the researchers gained access to the attacker’s stored instructions and working files, they were able to reconstruct the set-up of the attack step by step. According to ThreatLabz, the attacker used AI in virtually every phase. The full reconstruction is available in the ThreatLabz technical analysis.

Division of tasks between AI models

In the reconnaissance phase, the attacker assigned specific tasks to different AI systems. DeepSeek examined websites and applications for potential entry points. GLM then checked those results and looked for vulnerabilities that DeepSeek might have missed. Meanwhile, Grok carried out independent research from a separate server.

In the second phase, specialised AI agents were tasked with collecting and processing personal data, customer data and financial information. Among other things, they had to decrypt encrypted data, analyse transaction data and organise data files.

In the final phase, the attacker used AI to prepare the stolen data for sale. Data fields were translated into Chinese and superfluous information was removed. The attacker also asked AI where stolen data is typically sold.

Targets across multiple sectors

According to the researchers, the instructions show that the attacker had organisations in a wide range of sectors in their sights, including financial services, online retailers, education, logistics and dating platforms. The targets were located in countries including the United States, the United Kingdom, India, Japan, South Korea and several African nations. The main interest lay in employee and customer data and financial information.

ARTEX includes a default blocklist intended to restrict attacks on government bodies and universities. ThreatLabz says this restriction is relatively easy to circumvent.

Smaller players, bigger operations

According to the researchers, the key difference from manual attacks lies in the organisation. AI systems such as DeepSeek, GLM and Grok divide tasks, exchange results and continue research independently. As a result, a single attacker can set up an operation that previously required more people. ThreatLabz expects that smaller hacker collectives and individual cybercriminals will also be able to have a major impact in this way.

Albert Kramer, Solutions Consulting Manager Benelux at Zscaler, says the wide availability of AI tools such as Mythos and ARTEX is accelerating and increasing cyber risk. Kramer stresses the importance of installing security updates promptly and limiting access to applications, workloads and the connections between them through a Zero Trust approach.

The Dutch National Cyber Security Centre (NCSC) had already warned that AI can accelerate the exploitation of vulnerabilities. Zscaler also points to recent vulnerabilities in PHP as an example of the need to have security processes in order.

The full report can be read in the ThreatLabz technical analysis.

This article was originally published in Dutch on MSP Business. Read the original →