Anthropic claims to have blocked a cyber operation it says was largely executed by its own AI model after being manipulated by a China-backed group. The attack targeted dozens of global institutions.
The company said 30 entities, including financial firms and government bodies, were targeted in September. Several intrusions succeeded and allowed unauthorized access to internal data.
Anthropic said Claude Code performed up to 90% of the actions on its own. The attackers allegedly tricked Claude into believing it was conducting legitimate cybersecurity assessments.
Despite the autonomy of the system, Claude made frequent errors, fabricating details about targets and misrepresenting publicly available information as sensitive findings.
The cybersecurity community is divided. Some warn the incident shows how quickly AI capabilities are advancing, while others question whether Anthropic exaggerated the system’s independence for publicity.