A hacker has allegedly managed to steal a massive trove of sensitive data including tax and voter information from Mexican government sites after “jailbreaking” the Claude AI model of Anthropic.

Jailbreaking is a technique where a user manages to bypass safety filters and ethical guidelines of an Artificial Intelligence (AI) model through deceptive and persuasive prompts.

According to a Bloomberg report, the attack was flagged by cyber security researchers from Israel based Gambit Security , who reported that an unknown Claude user persuaded the chatbot to act as an “elite hacker” and find vulnerabilities in Mexican government networks, in order to exploit them and steal sensitive data. The report stated that initially Claude chatbot had refused to entertain pleas of the user, however, it complied after repeated persuasions through prompts in Spanish.

The attack started in December 2025 and continued for a month where the hacker allegedly used Spanish in prompts to jailbreak Claude. The Mexican government had also launched a parallel investigation into the cyberattack.

In total, 150 gigabytes of data from Mexican government network was stolen in the exploit, reported Bloomberg.

According to Gambit Security, the hacker breached Mexico’s federal tax authority and national electoral institute and multiple state governments’ sites, after persistently asking Claude for instructions.

The cybersecurity firm had informed Anthropic of the malicious actions of Claude user after which the company suspended the accounts. Anthropic has not released any official statement in the matter.

Also Read: Anthropic Accuses Chinese AI Companies of Illicit Distillation