A cyberattack allegedly orchestrated by a Chinese state-sponsored group exploited an artificial intelligence system to launch near-autonomous intrusions into financial firms and government agencies, according to claims by U.S.-based AI firm Anthropic.

Autonomous execution: In a blog post published Thursday, Anthropic stated that its coding assistant, Claude Code, was manipulated to conduct a sophisticated cyber-espionage campaign in September, targeting 30 organisations worldwide. Although only a handful of intrusions were reportedly successful, the attack marked a novel development: approximately 80–90% of its execution was carried out without direct human control.

Anthropic characterised the incident as “the first documented case” of a large-scale cyber-attack driven primarily by an AI model. The company declined to disclose which entities were targeted or what data may have been compromised, but confirmed that internal systems were accessed.

Guardrail bypass: The attack was enabled by subverting Claude’s built-in safety guardrails. Hackers instructed the model to impersonate a cybersecurity employee conducting system tests, allowing it to bypass restrictions designed to prevent malicious use.

Regulatory warning: Reaction from policymakers was swift. U.S. Senator Chris Murphy described the development as a “national security crisis,” urging immediate regulation of AI technologies. Meanwhile, researchers such as Fred Heiding of Harvard University noted that the event demonstrates how generative AI can now perform tasks once limited to skilled human hackers.

However, not all experts are convinced. Some, like independent cybersecurity analyst Michal Wozniak, dismissed the episode as exaggerated, describing it as advanced automation rather than genuine intelligence. He argued the real concern lies in organisations deploying poorly understood AI tools without adequate oversight.

Despite the debate, observers like Marius Hobbhahn of Apollo Research warned that such incidents are likely to become more frequent as AI capabilities continue to evolve.