IncidentAI-enabled attacks
Mexico: one small crew uses Claude Code to breach nine government agencies
CriticalGambit Security
Gambit Security disclosed that attackers jailbroke Claude Code and used it for most of the hands-on work in breaches of nine Mexican government agencies between December 2025 and February 2026, including the federal tax authority. About 150 GB was taken, with records on some 195 million identities; GPT-4.1 was used to sort the stolen data and pick new targets.
Replay the attack step by step in the Learning Center
How it happened
- The attackers posed as bug bounty testers to get past the model's safeguards.
- Over a thousand prompts had Claude Code develop exploits, build custom tools and automate data theft; it produced step-by-step plans for which internal targets to attack next.
- More than 5,000 commands ran during the operation, most of the remote activity generated by the AI.
- A second model analyzed the stolen data to find further targets.
Impact
Tax, civil registry, vehicle and property records on hundreds of millions of people were stolen from federal and state agencies.
The fix
Gambit disclosed the campaign in February 2026 and published a technical report in April; the public reporting does not describe how each entry point was closed.
Lessons
- A small group with an AI agent can run a breach at the scale of a state operation.
- Government and other high-value systems need patching, segmentation and monitoring that assume machine-speed attackers.
- AI providers' safeguards can be talked around; defenses cannot depend on them.