Anthropic just exposed the playbook. Russian spies, Chinese students, and lone hackers all used Claude to run autonomous cyberattacks.

Anthropic's September threat intelligence report just dropped the most detailed map of AI-powered cyber operations ever published.

One finding should stop every CISO cold:

A Russian-linked espionage group (GTG-20006) used Claude to automate their entire attack chain — reconnaissance, phishing, credential theft, lateral movement, data exfiltration.

When security products detected their malware, agents autonomously modified and redeployed the tools until they were undetected.

They targeted 20+ organizations: Ukrainian government, defense intel, drone manufacturers, European embassies.

But here's the structural shift that matters:

Chinese university students (GTG-10007) ran an autonomous vulnerability research program using Claude. They built an exploit foundry that found 12+ zero-days in a major security product in a single month.

A hacktivist using stolen API keys ran a multi-victim campaign that would have required a team of operators just one year ago.

AI collapsed the labor gap between state-sponsored operations and individual hackers. Sophistication is no longer a signal of who is behind an attack.

The report also found: attackers stole AI API keys from victims and used them as attack compute. The AI supply chain itself has become a criminal target.

Audit your agent permissions and API key exposure today. Treat AI credentials with the same rigor as production database credentials. The threat actor sitting across from you might be an autonomous agent running on someone else's stolen compute.

SOURCE: https://www.anthropic.com/threat-intelligence-report-september-2026
VERIFIED: Anthropic Threat Intelligence Report (Sep 10, 2026), Daniel Miessler analysis (Sep 15, 2026), The Guardian (Sep 12, 2026)
SIGNAL: This is the first comprehensive map of how AI agents are being weaponized across state espionage, financial crime, and hacktivism — and it proves the old threat-modeling playbook is dead.