Jiahao Shao, a computer science PhD student at the Hong Kong University of Science and Technology, says Anthropic restricted his paid Claude access after flagging his work as training a competitor. In a post on X on 31 October, he said he never distilled Claude: the rollouts came from a strong open-source model, and Claude only helped submit jobs and run analysis. He says anything touching RFT, SFT or training larger models now gets him blocked, his Opus 5.5 access downgraded to Opus 5, or the session killed outright.
Anthropic has not publicly commented on the account, so the case remains one user's account. The company's published rules are explicit that customers may not use its services or outputs to train competing models without written approval, a limit set out in its Commercial Terms, Consumer Terms and Usage Policy. In October Anthropic also detailed stronger anti-distillation defences: a two-stage classifier, a claimed false positive rate under 0.1 percent, and bans on tens of thousands of accounts, often within two hours of a campaign starting.
The post drew more than 3,600 likes and roughly 200,000 views, and the Hacker News thread on it passed 440 points with over 240 comments. Some commenters argued the blocks may have been triggered by region or VPN signals rather than training content, since Claude is not available in Hong Kong. For founders automating research or fine-tuning on top of Claude, the practical lesson is the same: keep training data and model outputs clearly separated, and expect classifiers to treat anything adjacent to model training as suspicious.