AI-assisted hackers target financial and personal data across 19 markets: Zscaler

• 26 Oct 2026
• TM Team
• 0 Comments

Zscaler's ThreatLabz researchers found that the attacker used a customized version of ARTEX, an AI-assisted penetration testing framework, alongside coding assistants and automation scripts. The investigation examined configuration files, scripts, and AI prompt histories recorded between August 4 and October 4, which showed attempts to obtain personal information, financial records, and other sensitive data.

According to Zscaler, the attacker modified the open-source ARTEX framework to support network scanning, vulnerability testing, and Android application analysis. The framework uses AI agents to break down security testing objectives into individual tasks and execute them using available tools.

The operator used AI coding environments, including Claude Code and Codex, alongside configurations specifying models from DeepSeek, GLM, and Grok. The recovered prompts instructed an AI model configured as DeepSeek to conduct initial investigations, with GLM assigned to review the findings for overlooked vulnerabilities. Other prompts assigned research tasks to Grok.

"One operator ran several AI models like a team: one researched, another checked the work, and scripts carried the context from one session to the next," Deepen Desai, executive vice president of cybersecurity at Zscaler, said.

Zscaler noted, however, that it could not independently verify which AI models processed every session. Some requests passed through third-party gateways, meaning the models specified in configuration files did not necessarily match those responding to the requests.

The attacker also configured scripts to retrieve saved findings, testing progress, and outstanding tasks whenever a new prompt was submitted. This allowed subsequent AI sessions to continue earlier investigations without repeating completed work.

Additional instructions assigned specialized AI agents to discover APIs, review source code, and examine Android applications. Some configurations permitted actions without individual operator approval, while prompts instructed the tools to treat assigned targets as already authorized for testing.

Zscaler identified prompts referencing organizations in 19 countries and territories, including Malaysia, Indonesia, Thailand, India, Japan, South Korea, the United Kingdom, and the United States. The records covered organizations across financial services, cryptocurrency, retail, education, logistics, and other industries.

"Malaysia is on the target list of a global AI-assisted hacking campaign," Desai said. "Malaysian organizations should treat themselves as in scope."

The attacker's objectives differed across industries. Prompts involving SMS providers sought administrative access and messaging credentials, while those targeting retail and educational services requested customer order records and purchase histories. Cryptocurrency-related instructions included attempts to obtain NFT gifts without payment.

Several prompts focused on obtaining personal information, particularly mobile phone numbers and transaction records. The operator also instructed AI tools to decrypt database fields, process transaction histories, and translate database structures into Chinese.

Other prompts sought information about marketplaces for selling stolen data. Zscaler assessed that the emphasis on personal and financial records indicated a financial motive, although the findings did not establish that every organization referenced in the prompts had been successfully compromised.

The operator also instructed AI assistants not to initiate direct withdrawals from financial accounts. According to Zscaler, the instructions cited concerns that unauthorized transactions would trigger security alerts or leave traceable records.