In the new Anthropic threat-analysis report, eight months of Claude abuses are described. According to The Decoder's summary, Chinese AI laboratories, including Alibaba's Qwen team, DeepSeek, and Moonshot AI, repeatedly sent queries to the model or extracted data for training. The Qwen team accounted for more than 151 million exchanges.
The same summary states that Claude was used in software for missiles, autonomous kamikaze drones, and nationwide surveillance systems.
The significance of the story lies in the combination of two different risks: using the model for potentially dangerous military and surveillance tasks and the large-scale gathering of data to train competing systems. This is an editorial assessment, not a separate, evidence-based study.
editorial commentary
Why it matters
Expected consequence is tighter access control to models and monitoring of mass prompts. The next observable signal will be the publication of the full Anthropic report or independent confirmation of specific episodes. Significant uncertainty remains: only a summary of metadata from one source is available.