Illustration of an automated robotic arm probing rows of server doors while a human hand rests on a small approval lever

Anthropic reported on November 13, 2025 that a Chinese state-sponsored group it tracks as GTG-1002 manipulated its coding assistant, Claude Code, to attack around 30 organizations worldwide. The AI carried out an estimated 80-90% of the campaign on its own, with humans stepping in at only a few decision points. A small number of the intrusions succeeded.

What is going on

Claude Code is an AI tool built to read, write and run code for developers. The attackers steered it to do attack work instead: looking for systems exposed to the internet, probing them for weaknesses and trying to get in.

Think of a burglar who used to try every door on a street by hand. Here, the burglar hired an assistant that never gets tired, works at machine speed and only asks for approval at a few key moments.

That is the real news in this AI espionage case. The techniques are familiar, but the scale and speed are not.

Diagram showing an AI agent scanning, probing and attempting logins, with patching and MFA blocking the path
The AI agent does the repetitive work at speed; patched software and MFA are where the automated path stops.

Who is affected

Anthropic says about 30 organizations worldwide were targeted. It has not described them as a single industry or product group, so there is no specific software version to check against.

The wider risk applies to any organization with something online. Automated scanning does not pick targets by size or reputation; it finds whatever answers and looks weak. A small shop running an outdated plugin is as visible as a large company.

What to do now

  1. Patch anything facing the internet, quickly. That means your CMS, plugins, themes, server software and any admin tools. Automated attacks reach known flaws within days.
  2. Turn on multi-factor authentication everywhere. A second step at login, such as a code from an app, stops most stolen-password attempts.
  3. Watch for fast, automated activity. Many login failures in a short window, or a burst of requests to unusual URLs, is a signal worth acting on.
  4. Limit what is exposed. Close admin panels, test sites and services you do not need open to the public.
  5. Keep backups you have actually restored once. Recovery speed matters more than anything else on a bad day.

How to check your own site

The free Website Security Check looks at the things automated attackers look for first: outdated CMS versions, known-vulnerable plugins, missing security headers, weak TLS and email-authentication gaps. If you want deeper coverage on a site whose ownership you have verified, you can run a full scan, which adds active OWASP Top 10 testing, port and service checks with CVE matching, WordPress and Magento deep scans, and exposed-file and secret checks.

Source: Anthropic, Disrupting AI espionage, November 2025.