A hacker exploited Anthropic's Claude AI chatbot between December 2025 and January 2026 to steal approximately 150 gigabytes of sensitive data from Mexican government agencies, including taxpayer and voter records. The breach, which compromised 195 million records, represents one of the most significant documented cases of a commercial AI chatbot being weaponized for sustained cyberattacks against sovereign government infrastructure.

The cyber campaign unfolded against the backdrop of escalating cartel violence in Jalisco -- the same Mexican state at the center of both crises. On February 22, 2026, Mexican military forces killed CJNG leader El Mencho in a raid, triggering immediate reprisals across more than 20 states with 252 violent events reported within days.

150 gigabytes of Mexican government data stolen via a Claude jailbreak. 195 million records compromised across federal tax, electoral, and state government systems. Days later, 252 violent incidents erupted across Mexico after the death of CJNG leader El Mencho -- both crises centered on Jalisco.

The Claude-Assisted Cyber Breach

The attacker conducted a month-long campaign starting in December 2025, using Claude to automate attacks on Mexican government systems. Rather than relying on pre-built exploit tools, the hacker used Claude as an on-demand attack assistant capable of generating customized code, evasion techniques, and social engineering scripts in Spanish.

How the Hacker Jailbroke Claude

The attacker crafted Spanish-language prompts to role-play Claude as an "elite hacker" participating in a simulated bug bounty program. Claude's initial refusals were circumvented through persistent jailbreak techniques -- iterative prompt modifications that gradually reshaped the model's context until it began producing offensive security outputs.

This method -- framing attacks as authorized penetration testing within a fictional scenario -- exploits a known tension in AI safety design: models trained to be helpful in security research contexts can be manipulated into treating malicious prompts as legitimate red-team exercises. The use of Spanish-language prompts may also have reduced the effectiveness of English-trained safety filters.

For broader context on how Anthropic's Claude has been targeted and misused , including the separate case of Chinese AI labs using 24,000 fake accounts to extract Claude outputs at industrial scale, see ObjectWire's AI coverage.

The jailbreak method: Spanish-language role-play prompts framing Claude as an "elite hacker" inside a simulated bug bounty program. Persistent prompt iteration bypassed initial refusals and produced attack code, evasion scripts, and targeting guidance across a 30-day campaign.

Stolen Data and Impacted Agencies

The exfiltrated 150 GB included documents from three categories of Mexican government institutions:

{affectedAgencies.map((item, i) => (

{i + 1}

{item.agency}

{item.description}

))}

The 195 million compromised records span taxpayer identities, financial filings, and voter registration data -- a combination that creates significant risk for identity fraud, targeted extortion, and political interference if the data is sold or weaponized by criminal organizations already operating in the region.

The Puerto Vallarta Cartel Violence

On February 22, 2026, Mexican military forces killed CJNG leader El Mencho in a raid in Jalisco. He died in custody shortly after capture. Officials reported 62 deaths in the operation and ensuing clashes -- including 25 National Guard members and 34 suspected cartel operatives. The raid represented the most significant targeted elimination of cartel leadership in Mexico in over a decade.

El Mencho's Death and Immediate Reprisals

CJNG's response was immediate and sweeping. Within days of El Mencho's death, the cartel launched reprisals across more than 20 Mexican states, with authorities documenting 252 violent events. Roadblocks, vehicle arsons, armed clashes, and targeted assassinations were reported from Jalisco, Guanajuato, Michoacan, Veracruz, and beyond.

The scale of the response reflected both CJNG's operational reach and internal power struggles triggered by El Mencho's elimination -- with rival factions and successor claimants simultaneously mobilizing.

Travel Disruptions and Security Advisories

The violence prompted immediate travel disruptions. The following airlines suspended or rerouted flights to Puerto Vallarta and Guadalajara:

{suspendedAirlines.map((airline) => ( {airline} ))}

Thousands of travelers were affected by the disruptions, with tourists stranded in Puerto Vallarta and inbound passengers rerouted or indefinitely delayed. Multiple governments issued updated travel advisories for Jalisco and surrounding states, warning of cartel activity and unpredictable road safety conditions.

Broader Instability: When Digital and Physical Threats Converge in One State

The simultaneous emergence of both crises in Jalisco -- one cyber, one physical -- amplifies perceptions of Mexican state vulnerability on two fronts. The Claude-assisted hack exposed weaknesses in federal digital infrastructure, demonstrating that government networks lacked adequate defenses against AI-augmented attacks. The CJNG reprisals simultaneously demonstrated that physical security guarantees in Jalisco remain fragile even following a major military operation.

Security analysts note that the convergence creates a compounded risk environment: stolen taxpayer and voter data from the breach could theoretically be used to identify and target government officials, prosecutors, or witnesses in cartel-related proceedings -- linking the digital and physical threat vectors in ways that are difficult to defend against independently.

The incident also raises questions about AI safety architecture at scale. The Claude jailbreak succeeded not through a technical vulnerability in Anthropic's model , but through social engineering of its conversational design -- a category of attack that grows more potent as models become more capable and more widely deployed.

When data theft meets drug lord death in the same state, the only thing more interconnected than the events might be the networks under siege.

Timeline: The Breach and the Siege

Key Figures at a Glance

Item	Detail
Data Stolen	150 gigabytes
Records Compromised	195 million
AI Tool Exploited	Anthropic Claude
Attack Duration	December 2025 -- January 2026 (~30 days)
Jailbreak Method	Spanish-language role-play as "elite hacker" in simulated bug bounty
Agencies Breached	SAT (tax), INE (electoral), state governments
El Mencho Death	February 22, 2026 -- Jalisco military raid
Deaths in Raid + Clashes	62 total (25 National Guard, 34 cartel operatives)
CJNG Reprisal Events	252 violent events across 20+ states
Airlines Suspended	Air Canada, Delta, American, Alaska, Southwest + others