Hacker exploits AI tools to breach 9 Mexican government agencies | brief

Nine Mexican government agencies were compromised by a single hacker who exploited two popular AI platforms, Claude Code and OpenAI's GPT-4.1, between December 2025 and February 2026. Researchers from Gambit Security observed the attacker operating at a speed that outpaced human security teams, with Claude Code executing approximately 75% of the remote commands sent to government computers, as reported by HackRead.

The attacker manipulated the AI platforms by initially posing as part of a bug bounty program and providing a hacking manual to bypass safety filters. A custom tool, BACKUPOSINT.py, was used to exfiltrate data from 305 internal servers, which GPT-4 then analyzed to create detailed reports on government server configurations. The breaches affected federal and state systems, including the federal tax authority (SAT), where 195 million taxpayer records were accessed, and Mexico City, where 220 million civil records were compromised. The hacker also gained control of a server system in Jalisco state, accessing sensitive health and domestic violence victim data.

The attacker utilized 20 custom scripts targeting 20 CVEs and rephrased commands when the AI initially refused requests, demonstrating a blend of basic exploitation techniques with advanced AI tools. The success of the attack was attributed to outdated government security practices, such as infrequent software updates and password changes.