25,000 records exposed as AI agents automate bank thefts

At least 25,000 personal credit records from Shinhan Bank customers were leaked in South Korea, a breach that occurred just days before security firm Straiker released findings on a similar global threat. These new leaks provide a concrete measure of the data exposure resulting from the automated cyberattacks that have targeted the region's financial sector.

Straiker’s research identifies the hacking group CyberCartel as the primary actor behind a new wave of automated assaults on Latin American banks and mobile payment apps. Active since at least 2012, the group has historically relied on human labor to execute thefts. However, Straiker indicates that the integration of artificial intelligence has allowed the group to expand its reach and increase its pace of attacks far beyond what manual efforts could achieve. In recent months, the group has successfully stolen passwords and customer data, accessing bank accounts with minimal direct intervention from the criminals.

The operation utilizes specific AI agents, including Grok from SpaceX, Claude from Anthropic, and a model from the Chinese lab Moonshot AI. Straiker shared its findings with SpaceX and Anthropic, both of which are taking steps to restrict the group’s access. Anthropic has already blocked CyberCartel’s use of Claude, while Straiker is still attempting to contact Moonshot AI.

The attacks succeed by exploiting a known weakness in AI safety protocols. The models are typically programmed to refuse requests that aid in cyberattacks, but CyberCartel bypasses these safeguards by instructing the agents that they are operating in test environments. If one model refuses to proceed, the hackers switch to another. If that model also declines, the agents are assigned to deploy malware instead. This method demonstrates the fragility of the guardrails currently built into these systems.

This incident is not the first time Anthropic has faced such challenges. In November 2025, the company disclosed that its AI was used in a cyberattack. Later, Claude was used by researchers to breach the security of rival OpenAI, not criminals. Anthropic also warned earlier this year that its new Mythos model could be exploited for similar attacks, prompting security teams to fix thousands of vulnerabilities. Straiker’s current report does not suggest that CyberCartel has used the Mythos model.

The timing of the Straiker report coincides with investigations by South Korean authorities into a comparable problem affecting Hana Bank, KB Kookmin Bank, and Shinhan Bank. The leak of 25,000 customer records from Shinhan Bank underscores the immediate impact of these automated methods. Industry players are responding to the threat; Nvidia recently launched the Open Agent Safety Platform, a system designed to contain rogue agents and prevent them from escaping sandboxes or executing unauthorized code.