A report citing The Information relays anonymous sources’ allegations that DeepSeek and Moonshot used OpenAI Astra outputs to train models, naming DeepSeek V4.1 Flash 1017 and Kimi K3.1 Pro. It describes a shift in targeting after Anthropic tightened defenses against output collection. Anthropic’s claimed 90–99% reduction is a company assessment, not an independently verified result. OpenAI says it continually adjusts its defenses; DeepSeek did not respond and Moonshot declined comment. The publicly described evidence does not establish each model’s training provenance or resolve whether a violation occurred.
Back to Tech
Report alleges Chinese AI labs trained on Astra outputs
Anonymous-source allegations and provider defense claims are not independently established findings.