Hacker Used Anthropic's Claude to Steal Sensitive Mexican Data
An attacker spent about a month from December 2025 using Claude against Mexican government systems, including the tax authority and the electoral institute, and took roughly 150 gigabytes: records on some 195 million taxpayers, voter information, civil registry files and staff credentials. The route past the model's refusals was not technical. Researchers at Gambit Security, who found it, describe Spanish-language prompts asking Claude to role-play as an elite hacker, repeated until it complied. Anthropic banned the accounts and strengthened its safeguards. Read beside GTG-1002 three months earlier: that was a state-sponsored group with an attack framework around the model, and this is one person with persistence and a second language, which is the more repeatable of the two.
