DeepSeek and Moonshot AI Accused of Copying GPT-6 Astra Capabilities Through Distillation Attacks

The Information reported, citing people with direct knowledge, that Chinese AI start-ups DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's GPT-6 Astra to train their latest models, Moonshot's Kimi K3.1 Pro and DeepSeek's upgraded V4.1-Flash 1017, in alleged model-distillation campaigns. Anthropic said early evidence suggested its new safeguards cut data harvested by distillers by around 90% to 99% and reduced reasoning tokens extracted on its platform by 10 to 100 times, banning tens of thousands of accounts, and sources said those restrictions pushed the labs toward OpenAI's relatively looser API defences. No server logs, account records or training data have been published to substantiate the Astra allegations, and DeepSeek, Moonshot AI and OpenAI had not publicly confirmed them at the time of publication.

DeepSeek and Moonshot AI Accused of Copying GPT-6 Astra Capabilities Through Distillation Attacks The allegations, based largely on anonymous sources and statements from US AI providers, remain unverified by publicly available technical evidence Chinese artificial intelligence start-ups DeepSeek and Moonshot AI have reportedly targeted OpenAI's flagship model, GPT-6 Astra https://www.ibtimes.co.uk/gpt-6-astra-ran-human-starcraft-bot-1823581 , in campaigns designed to extract its capabilities through a technique known as model distillation. The Information reported, citing people with direct knowledge, that DeepSeek and Moonshot AI relied heavily on outputs from GPT-6 Astra to train their latest models. Other sources subsequently reported the findings, citing The Information. The models named in the report were Moonshot's Kimi K3.1 Pro and DeepSeek's upgraded V4.1-Flash 1017. The alleged campaigns followed tighter security measures by Anthropic around Claude's reasoning process, which reportedly made its models harder to probe. DeepSeek, Moonshot AI and OpenAI had not publicly confirmed the specific Astra allegations at the time of publication. Anthropic's Tighter Defences Reportedly Shifted Attention Towards OpenAI Anthropic said early evidence suggested its new safeguards https://www.ibtimes.co.uk/anthropic-ai-models-safety-research-warning-1828579 had reduced the amount of data harvested by distillers by around 90% to 99%. It also reported that reasoning tokens extracted on its platform had fallen by 10 to 100 times following the introduction of its new classifiers. The company said it had banned tens of thousands of accounts associated with distillation, often within two hours of a campaign starting. Sources cited in the reporting said the restrictions pushed the Chinese labs towards OpenAI, whose API defences were considered relatively looser. Distillation involves training a smaller model, known as a student, using the outputs of a more powerful teacher model. This allows developers to reproduce some of the larger model's capabilities without the expense of building an equivalent system from scratch. OpenAI and Anthropic https://www.ibtimes.co.uk/openai-anthropic-ai-safety-incidents-1822208 have separately reported that some distillation campaigns used proxy networks https://www.ibtimes.co.uk/alibaba-ai-model-secretly-tried-send-data-beijing-server-shock-test-rival-cohere-claims-1828481 , fraudulent accounts and automated prompts to extract protected reasoning traces. These techniques allow attackers to collect information about how a model approaches problems, rather than just its final answers. However, the available reporting does not establish that every technique was used in the specific Astra-related campaigns. OpenAI And Anthropic Have Previously Reported Similar Campaigns OpenAI said on 30 September 2026 that it had disrupted a coordinated campaign dating back to July. The operation allegedly sought to extract hidden reasoning pathways from its models. The company identified related activity accross more than 15,000 users and attributed a core cluster to individuals linked to Moonshot AI. Moonshot has previously denied developing its models through distillation. Anthropic has also reported detecting campaigns involving seven Chinese AI labs since February 2026. Its initial February disclosure concerned three labs and involved more than 16 million exchanges accross approximately 24,000 accounts. Evidence Behind the Latest Allegations Remains Unpublished The latest claims rely largely on anonymous sources and statements from US API providers. No independent publication of server logs, account records or training data has established the alleged campaigns. Similarities between model outputs alone do not establish that one company used another's responses to train its models. The specific Astra allegations remain unverified by publicly available technical records. The allegations primarily concern potential breaches of service agreements rather than conduct established through court proceedings. OpenAI's Services Agreement generally prohibits customers from using its outputs to develop AI models https://www.ibtimes.co.uk/openai-ai-deepfake-evidence-courts-1826007 that compete with its products and services, subject to specified exceptions. OpenAI has also warned that extracting protected reasoning traces could allow competitors to reproduce advanced problem-solving methods while bypassing safety filters built into the original models. © Copyright IBTimes 2026. All rights reserved.