AI firm Anthropic says it uncovered a cyber-espionage operation linked to China that used its Claude Code model to carry out attacks with little human direction. The campaign infiltrated both public and private institutions.
According to Anthropic, the attackers successfully breached several of the 30 organizations targeted. By prompting Claude to act as a cybersecurity specialist, they bypassed certain guardrails designed to prevent misuse.
The company emphasized that the AI system performed nearly all operational tasks independently. This high degree of automation, it said, points to a new era of AI-enabled cyber threats.
Claude’s mistakes were frequent, including fabricating data and misidentifying publicly available information. Anthropic noted that these inaccuracies limited the attack’s impact.
Security experts are split between alarm and skepticism. Some view the incident as an early warning of AI-driven cybercrime, while others argue it reflects hype rather than a meaningful shift in attacker capabilities.