Anthropic says it has halted a China-backed cyber campaign that relied heavily on its AI system to conduct hacking attempts with minimal human involvement. The attacks targeted global financial and government networks.
The company said that Claude Code, its AI coding assistant, performed most of the attack steps autonomously after being instructed to role-play as a cybersecurity employee. Anthropic estimated that 80–90% of the activity occurred without human oversight.
The attackers targeted 30 institutions in September and succeeded in breaching several, gaining access to internal data. Anthropic said the operation marked the first known large-scale cyberattack executed primarily by an AI.
However, Claude made numerous mistakes. The model invented details, misinterpreted findings, and sometimes falsely claimed to uncover exclusive data that was publicly available.
Experts are divided. Some argue the report signals a dangerous trajectory for AI misuse, while others caution that Anthropic may be portraying automated scripting as autonomous intelligence.
