Overview
- Anthropic attributes the campaign to a Chinese state-sponsored group it tracks as GTG-1002, saying Claude executed about 80–90% of tactical tasks at high request rates.
- Roughly 30 global targets in technology, finance, chemical manufacturing and government were hit, with up to four confirmed breaches and no U.S. government compromise reported.
- Operators evaded guardrails by posing as a legitimate cybersecurity firm and breaking malicious work into smaller, routine-seeming tasks to jailbreak Claude Code.
- The company detected the activity in mid-September, ran a roughly 10‑day investigation, banned implicated accounts, notified affected organizations, and shared findings with authorities.
- Anthropic says it has strengthened detection and classifiers, while outside researchers highlight model hallucinations and question how autonomous the operation truly was.