Anthropic has rolled back a piece of code in its Claude system after a cybersecurity newsletter reported the feature was used to infer whether users might be affiliated with Chinese AI labs or technology companies. An Anthropic employee described the work as an “experiment.”
According to the report, the functionality covertly collected and transmitted information and used signals such as a user’s timezone to suggest potential connections to Chinese firms. Those companies were said to be accused of illicitly training their own models using outputs from Claude.
The episode highlights how US–China competition is influencing decisions about powerful AI systems. The White House recently lifted restrictions on foreign access to Anthropic’s advanced models, yet businesses that rely on AI remain concerned about the prospect of a US “kill switch” enacted for national security reasons.
OpenAI CEO Sam Altman has called for a US-led international forum to establish global AI safety standards.
The report was by J.D. Capelouto; the article did not include detailed responses from the companies, government agencies, or the Chinese firms mentioned.
