Claude used to attack Mexican agencies, researchers say
Researchers say one attacker used AI to break into Mexican government agencies, and some agencies deny it. Either way, the basics now matter more.
Published · 3 min read

On 25 February 2026, Gambit Security, an Israeli security firm, published research saying that an unidentified attacker had used Anthropic's Claude to break into Mexican government agencies from December, taking 150 GB of data, including tax and voter information, Bloomberg reported. According to the research, the attacker wrote prompts in Spanish, told the AI to follow bug bounty rules, and found ways around it when it objected. Anthropic said it had investigated, disrupted the activity and banned the accounts involved, and some of the agencies dispute the research: according to Engadget, Jalisco's state government denied being breached, saying only federal networks were affected, and the national electoral institute denied any breach or unauthorised access. Mexico's tax authority, the SAT, said it had found no illegitimate access or unusual activity in its systems. Whoever turns out to be right about each agency, the researchers' account of what the AI changed is worth every office's attention.
What the researchers say the AI changed
Gambit's full report, published on 10 April, says the campaign ran from late December 2025 to mid-February 2026 and breached government bodies at federal, state and municipal level. By its count, about three quarters of the commands run remotely on the compromised servers were generated and executed by Claude Code. Gambit calls this a significant evolution in offensive capability, and much of what it documents is speed: the AI "compressed attack timelines to below standard detection and response windows", and let one operator handle volumes that would normally need a team of analysts. "The attacker's questions revealed his knowledge gaps," the report says of one stage.
The ways in, Gambit says, were old ones
Gambit counts 20 exploits aimed at 20 different published vulnerabilities, reports evidence that some of the systems were end-of-life or out of support and so could not receive security updates, and lists about 2,000 user accounts at one state government compromised along with their plain-text passwords. Many of the vulnerabilities, it concludes, "were addressable through standard security controls", which it names as patching, credential rotation, network segmentation and endpoint detection.
That is the lesson for an office here, whether or not every agency was breached. The same tools are open to anyone, and what they lower is the skill needed to turn a known, unfixed weakness into a break-in, and the time it takes. An attacker who once needed a team, or years of practice, to chain several weaknesses together can now ask for help at each step. The basics were always the best defence. They are now also the most urgent.
What it means for the system nobody will touch
Many offices have one: the old database behind a portal, the server under a desk, the application whose builder left years ago. A system nobody wants to touch is usually a system nobody patches, and one past its end of support cannot be patched at all, which makes moving it a security decision as much as a practical one. That is the work behind Move it, which starts with a two-week trial run on a copy, so the live data is never touched while the way off the old system is worked out.
Five checks that matter more now
- Apply published security fixes on a schedule measured in days, starting with anything reachable from the internet.
- List the systems that can no longer be patched, such as end-of-life software or the old server nobody owns, and set a date to replace or isolate each one.
- Check how passwords are stored. They should be hashed, never readable as text, and default or shared accounts should be gone.
- Rotate passwords and keys when someone with access leaves, and after any suspected break-in.
- Separate systems on the network, so a break-in on one server does not reach the database behind another, and set alerts for bulk exports that reach a person the same day.
If one of your systems is the one nobody will touch, the fit check is where I would start.
Sources
- Gambit Security, The AI-Assisted Breach of Mexico's Government Infrastructure
- Gambit Security, A Single Operator, Two AI Platforms, Nine Government Agencies: The Full Technical Report, 10 April 2026
- Bloomberg Law, Hacker Used Anthropic's Claude to Steal Mexican Data Trove (2), 26 February 2026
- Engadget, Hacker used Anthropic's Claude chatbot to attack multiple government agencies in Mexico, 25 February 2026
- Expansión, El SAT niega hackeo a sus sistema para robar datos de contribuyentes y habla de su ciberseguridad, 26 February 2026
Enrique Lacambra III builds internal business systems from Cainta, Rizal.
Tell me what is broken and I will tell you if it is worth building. A 20-minute call, no charge, and I will say so if you do not need me.
