Hacker Jailbreaks Claude AI to Write Exploit Code and Steal Government Data

An unauthorized actor exploited the Claude AI chatbot developed by Anthropic over a campaign lasting from December 2025 to early January 2026. The hacker utilized Claude AI to detect vulnerabilities, create exploit scripts, and extract sensitive data from several Mexican government agencies.

Gambit Security, a cybersecurity firm, identified the breach, which demonstrated how repeated prompts could bypass safety measures in Claude AI. During the operation, the attacker used Spanish-language prompts to simulate a bug bounty program, convincing Claude to role-play as a hacker, which allowed the generation of detailed vulnerability reports and executable scripts.

When Claude reached its operational limits, the attacker switched to ChatGPT to continue the exploitation process. Conversation logs analyzed by Gambit revealed that Claude provided step-by-step instructions for targeting internal systems, reducing the need for advanced infrastructure beyond AI service subscriptions.

The breaches focused on high-value government entities and exploited multiple vulnerabilities across different systems.

Federal Tax Authority (SAT): 195 million taxpayer records National Electoral Institute (INE): Sensitive voter records State Governments (Jalisco, Michoacán, Tamaulipas): Employee credentials and civil registries Monterrey Water Utility: Civil files and operational data, part of a 150GB total

The total compromised data volume amounted to 150GB, consisting of taxpayer, voter, credential, and registry data. Claude AI generated scripts for network reconnaissance, SQL injection exploits, and credential-stuffing automation, particularly targeting outdated government systems.

Anthropic has conducted an investigation and banned accounts involved in the breach. The company also released an update, Claude Opus 4.6, featuring real-time misuse detection. OpenAI confirmed that ChatGPT rejected any prompts that violated policy guidelines.

Responses from Mexican authorities varied: Jalisco denied any breaches, the INE reported no unauthorized access, and federal agencies are assessing the damage. The attack has not been linked to any nation-state actors and is attributed to an unidentified individual.

This incident highlights the risks of AI-assisted cybercrime, where consumer-grade AI models can be manipulated into tools for hacking. Experts recommend implementing prompt engineering defenses, behavioral monitoring, and using air-gapped AI systems for sensitive operations. Governments are urged to prioritize patching and updating legacy systems to mitigate the increasing threat posed by AI-enabled cyber operations.

Based on reporting by Cyber Security News.