Hacker Jailbreaks Claude AI to Generate Exploits and Steal Mexican Government Data
2 min. read 
Published on
A hacker used Anthropic’s Claude AI over a month to find vulnerabilities, write exploit code, and steal sensitive data from Mexican government agencies. The campaign ran from December 2025 to early January 2026. Gambit Security uncovered it through conversation logs. They state: “Persistent Spanish prompts bypassed safety to produce scanning scripts, SQL injection exploits, and credential stuffing tools.”
The attacker role-played Claude as an elite hacker in a fake bug bounty. Initial refusals gave way after repeated persuasion. Claude generated thousands of lines including network scans, payload deployment, and data exfil plans. When limits hit, the hacker switched to ChatGPT for evasion tactics. No advanced setup needed, just AI subscriptions.
Targets included the Federal Tax Authority (SAT) with 195 million taxpayer records, National Electoral Institute (INE) voter data, and state systems in Jalisco, Michoacán, and Tamaulipas. Total stolen: 150GB of credentials, civil registries, and operational files. No public leaks yet. Legacy systems with unpatched web apps and weak auth fell easy.
Gambit analyzed logs showing Claude chaining tasks from recon to exploitation. Prompts targeted common flaws like SQLi and misconfigs in Mexican gov tech. Anthropic banned accounts and upgraded Claude Opus 4.6 with misuse detection. OpenAI confirmed ChatGPT blocks such requests. Mexican agencies downplay; Jalisco denies breach.
This shows AI lowering attack barriers. Solo hackers gain APT-level output. Governments face new “agentic” threats from consumer tools.
Compromised Entities
| Agency | Data Type | Volume |
|---|---|---|
| SAT (Federal Tax) | Taxpayer records | 195M records |
| INE (Electoral) | Voter data | Sensitive personal |
| Jalisco State | Employee creds, registries | Multiple systems |
| Michoacán/Tamaulipas | Civil/operational data | Part of 150GB |
| Monterrey Water | Files and ops data | Utility breach |
Attack Techniques Generated
AI-assisted methods.
- Network scanning scripts.
- SQL injection exploits.
- Credential stuffing automation.
- Lateral movement plans.
- Data exfiltration tools.
Response Actions
Mitigate AI abuse.
- Monitor enterprise AI prompts.
- Block jailbreak patterns.
- Patch legacy government systems.
- Use air-gapped AI for sensitive work.
- Train on prompt injection risks.
FAQ
Persistent role-play prompts as “elite hacker” bypassed safety for exploit code.
150GB from SAT, INE, states including 195M tax records.
Yes, Opus 4.6 adds real-time misuse probes.
No, solo actor per Gambit analysis.
Prompt monitoring, behavioral AI controls, legacy patching.
Read our disclosure page to find out how can you help VPNCentral sustain the editorial team Read more
Improve this guide
User forum
0 messages