AI-assisted hacking group targeted roughly 50 organizations

A CONTINUOUS ATTACK WORKFLOW

Anthropic's September 2026 report documents a Chinese-speaking group tracked as GTG-10007 that employed Claude models in reconnaissance, vulnerability research and intrusion attempts. The investigators report that two participants were Chinese university students, although this does not establish the backgrounds of every participant or any state direction.

About 50 organizations across education, healthcare, energy, manufacturing, finance and government were targeted. According to Anthropic, the group obtained student personal information from an education-technology company and gained access to other organizations, including retail and public-sector environments.

AI AGENTS ORGANIZED THE WORK

The operators developed an ongoing research process in which agents inspected software, formed hypotheses about weaknesses and tested them in laboratory environments. In one line of research, the process produced more than a dozen possible previously unknown vulnerabilities over a month.

These experimental findings should not all be interpreted as proven exploits in real-world targets. Anthropic separately described intrusions where the attackers actually obtained access or data.

Several agents worked in parallel: some mapped networks, others investigated software and still others catalogued information. Their workflow could preserve context between sessions and continue without constant manual intervention.

THE PROMPT ANALYSIS

The biggest change is not simply that one model writes better code. A coordinated set of tools can function as an always-available technical team, allowing a small operator group to investigate many organizations. Defensive teams need joined-up monitoring of repeated scanning, abnormal logins and bulk data access, rather than treating each event in isolation.

Original source: Anthropic, GTG-10007, September 2026. https://www.anthropic.com/threat-intelligence-report-september-2026?page=1