Anthropic says it has stopped what it calls the first major cyberattack largely executed by an AI model after detecting misuse of its Claude Code system by a Chinese-linked hacking group. The attack affected institutions across multiple sectors.
The campaign targeted 30 organizations, and several breaches were confirmed, with attackers gaining access to internal data. Anthropic says the hackers bypassed safeguards by instructing Claude to impersonate a cybersecurity professional.
The company said the AI model autonomously completed 80–90% of the operational workflow. It claims this level of automation marks a turning point in cyber threat evolution.
However, the AI displayed multiple weaknesses. Claude fabricated information, misunderstood target data, and incorrectly labeled open-source information as sensitive.
Experts question how much autonomy the AI truly exercised. Some fear the implications, while others say the episode appears exaggerated and closer to automated scripting than genuine independent action.
