AI Agents Automate Bank Heists in Latin America

The cost of cybersecurity is rising as criminal organizations adopt technology that requires less manual labor to execute high-value thefts. A long-standing hacking group called CyberCartel has begun using artificial intelligence agents to streamline attacks on banks and mobile payment applications across Latin America. According to new research from the cybersecurity startup Straiker, this shift has dramatically expanded the group's reach. The automation allows the attackers to target more organizations simultaneously while moving at a pace that human-driven hacking simply cannot match. This development signals a significant change in how financial institutions must defend their digital infrastructure against automated threats.

CyberCartel has been active since at least 2012, making it a well-established presence in the region's security landscape. However, the recent integration of AI tools has marked a distinct acceleration in its capabilities. The group now steals passwords and customer data with minimal direct involvement from the criminals themselves. Individual bank accounts are being accessed and drained through automated processes rather than manual exploitation. This efficiency reduces the time between initial breach and financial loss, putting pressure on banks to detect and respond to these automated intrusions in near real-time.

The specific tools driving this automation include agents from several major AI developers. CyberCartel has utilized Grok from SpaceX and Claude from Anthropic. Additionally, an agent from the Chinese lab Moonshot AI has been part of the operation. Straiker shared its findings with both SpaceX and Anthropic. The two companies are currently working to cut the group off from their services to prevent further exploitation. Anthropic has already blocked the group's access to Claude. Straiker is still attempting to reach Moonshot AI to discuss the breach and potential mitigations.

The incident highlights the fragility of safeguards built into modern AI models. These systems are normally designed to refuse requests that assist in cyberattacks. However, CyberCartel reportedly bypassed these defenses by telling the models they were operating in test environments. When one model declined to continue, the hackers switched to another. If that model also refused, the AI agents were assigned to deploy malware instead. This adaptive approach allows attackers to work around standard safety protocols without needing to develop entirely new exploits for each model.

Straiker's research does not suggest that the group has used Anthropic's Mythos model. Anthropic drew attention earlier this year when it unveiled Mythos and warned that hackers could exploit it for exactly this kind of attack. The release of the model prompted security teams around the world to rush to fix thousands of vulnerabilities that the model had identified. This proactive warning underscored the dual nature of advanced AI, which can both secure systems and identify weaknesses for exploitation by bad actors.

This is not the first time Anthropic has learned that malicious actors are using its technology. In November 2025, the company disclosed details of its AI being used in a cyberattack. Later, Claude was used to break the security of OpenAI, a rival AI developer. That specific operation was carried out by researchers rather than criminals, but it demonstrated the power of AI in breaking down defensive layers. The current situation with CyberCartel shows that criminal groups are rapidly adopting similar techniques for their own gain.

The Straiker report arrives just days after word emerged that South Korean authorities are investigating a comparable problem affecting banks in that country. Hana Bank, KB Kookmin Bank, and Shinhan Bank are among those affected. The personal credit information of at least 25,000 Shinhan Bank customers has been leaked. This parallel development suggests that the use of AI in banking attacks is not an isolated incident but a growing trend across different regions and financial systems.

Companies in the industry are aware of the risks posed by rogue agents. Nvidia, now focused on hardware for AI data centers, recently launched the Open Agent Safety Platform. This is a software platform and reference system design that uses a new hardware and software security stack to contain rogue AI agents. It is intended to help stop agents from escaping their sandboxes, executing unauthorized code, gaining unauthorized access to critical infrastructure, or bypassing guardrails. Such solutions are becoming critical as the line between helpful automation and automated intrusion grows thinner.

The next steps for financial institutions will likely involve tighter controls on AI agent permissions and more robust monitoring of their behavior. Banks must decide how much autonomy to grant these systems while ensuring they cannot act beyond their intended scope. The race to secure AI-driven finance is just beginning, and the stakes are high for both providers and consumers who rely on the stability of their payment apps and bank accounts.