tl;dr: Researcher Jiahao Shao claims Anthropic flagged him for "training a competitor" despite only using Claude to submit jobs and run analysis on rollouts from an open-source model. He reports that any prompts touching RFT, SFT, or training larger models now trigger blocks, model downgrades from Opus 5.5 to Opus 5, or session termination. He criticizes Anthropic for treating paying research users as adversaries while having trained on public data themselves.