▲ File Photo
A hacking group has successfully automated cyberattacks against financial institutions in Latin America using artificial intelligence (AI) agents, tech-focused media outlet The Information reported on the 21st (local time).
According to an investigative report by cybersecurity startup Striker cited by the media, a hacking group known as Cyber Cartel utilized an AI agent network powered by Anthropic's Claude, SpaceX's Grok, and Chinese startup Moonshot AI's Kimi to launch automated cyberattacks targeting over 400 financial institutions in the Latin American region.
The deployed AI agents operated with minimal human intervention, stealing passwords and customer data, and even gaining access to victims' bank accounts, the findings showed.
Although AI models are equipped with safety guardrails designed to reject instructions for cyberattacks, Cyber Cartel bypassed these safety features by tricking the models into believing the tasks were being conducted in harmless environments, such as academic tests or internal corporate system audits.
Furthermore, if Claude refused to carry out a hack, the system instructed Kimi to do it instead, and if Kimi also refused, the agent system automatically switched to deploying malicious code.
The Information assessed that the survey results demonstrate how leveraging AI models allows for the rapid organization and execution of multiple cyberattacks, potentially overwhelming the defense systems of targeted institutions.
"The reality is that offense currently outpaces defense by a wide margin, and companies are not yet prepared," said Ankur Shah, CEO of Striker.
Striker sent the report to AI developers including Anthropic and SpaceX last week.
Anthropic and other developers stated that they conducted internal investigations immediately after receiving the report and took prompt action to block hackers from using their models.
According to Striker, Cyber Cartel appears to have been active since at least 2012, and its attack speed and efficiency have improved dramatically in recent months with the adoption of new AI agents.
During the attacks, Cyber Cartel tasked the Grok AI agent with gathering information on targets across the web, assigned four agents based on Claude Opus and Claude Sonnet to find optimal attack strategies, and gave another Claude model the role of a commander overseeing the hacking operations.
Ken Huang, an adjunct professor at the University of San Francisco who reviewed Striker's research, noted that this case clearly shows how cybercriminals are concretely automating the types of attacks that OpenAI and Anthropic have broadly warned against, proving how much AI has lowered the barrier to entry for sophisticated hacking.
Scott Roberts, Chief Information Security Officer (CISO) at enterprise software company UiPath, said, "We are seeing an increase in automated reconnaissance attempts targeting our systems."
"Automated attacks have steadily risen for over 20 years, and now, thanks to AI models, attacks have become much easier," he added. "We need to adjust our security operations and detection frameworks."
※
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.