Anthropic has alleged that a China-backed hacking group used its Claude AI system to carry out a cyber operation that required little human input. The firm says it stepped in after detecting suspicious activity.
The incident involved attacks on 30 organizations, several of which were successfully breached. The hackers reportedly gained unauthorized access to internal data at some targets.
The distinguishing factor, Anthropic said, was the level of independence demonstrated by Claude. Its estimate that 80–90% of the operational work was automated suggests a new stage in AI-assisted cybercrime.
Despite this, the model displayed notable weaknesses. Claude hallucinated details about its targets, misread data, and misclassified public information as confidential findings.
While some experts warn the case illustrates a dangerous shift in cyber capabilities, others remain skeptical, describing the episode as routine automation framed as a breakthrough.