Anthropic is blocking paying users over AI model fine-tuning attempts

Anthropic has faced criticism after its automated security systems began restricting access for paying users. The blocks and access downgrades were triggered by workflows involving attempts to automate data preparation or fine-tune large language models using the platform.

• A PhD student from Hong Kong stated that his post about the restrictions on X gained over 277,000 views.
• Anthropic's security system interpreted pipeline automation as a preparation for building a competing model.
• The researcher reported forced session terminations, version downgrades, and blocks during fine-tuning attempts.

The Incident with the Hong Kong Researcher

Jiahao Shao, a PhD student at the Hong Kong University of Science and Technology, reported facing severe limitations while using Claude. According to him, he used a strong open-source model to generate training examples while employing Claude to automate the process, send tasks, and analyze results. However, Anthropic's security system flagged these actions as suspicious.

Shao stated that his active sessions were abruptly terminated, his access level was downgraded from Opus 5.5 to Opus 5, and any operations related to fine-tuning were blocked. The researcher publicly appealed to Anthropic CEO Dario Amodei, criticizing what he perceived as double standards regarding data usage.

Technical Details and Fine-Tuning Methods

The restrictions affected workflows involving supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). The first approach relies on training models using pre-prepared correct examples, while the second utilizes a reward mechanism to optimize and improve model performance for specific tasks.

Both methods are standard practices when adapting artificial intelligence to specialized research or commercial applications. Nevertheless, Anthropic's monitoring algorithms seemingly interpreted the automated coordination of these processes as an attempt to synthesize data for a rival system.

Boundaries of API Usage and Developer Concerns

This case highlights a critical question for software developers: where is the exact boundary between legitimate API usage and actions that a model provider considers an attempt to build a competing system? Experiment automation and synthetic data generation remain vital parts of modern research pipelines.

At the time of the report, Anthropic had not publicly commented on the incident. Without an official explanation from the company, it remains unclear which specific actions triggered the automated restrictions, sparking a broader debate on how transparent AI platform rules should be.

What This Means for the Tech Industry

The situation demonstrates the vulnerability of independent researchers relying on closed AI lab ecosystems. When paying users face sudden restrictions without clear explanations, it undermines trust in cloud platforms and forces teams to look for more predictable or open alternatives.