Anthropic blocks researcher for using Claude to automate model training
Anthropic is killing sessions and downgrading users who use Claude to automate research for competing models.
Drama · Source: Hacker News
What happened
Anthropic blocked computer science researcher Jiahao Shao. The company accused him of training a competitor. Shao says he never distilled Claude to copy its capabilities. He only used the model to submit jobs and run analysis for an open-source project. Anthropic flagged the activity anyway. The company now treats paying users doing basic research automation as a threat. Founders rely on these tools to move fast. Having your account restricted for running basic analysis is a massive roadblock.
The restrictions are severe and automated. Shao reports that any prompt touching supervised fine-tuning, reinforcement fine-tuning, or training larger models triggers a harsh penalty. The system does not ask for clarification. It either blocks the user entirely, kills the active session outright, or downgrades them from Opus 5.5 to the older Opus 5 model. Anthropic is actively policing user prompts to detect any hint of AI development. This is a draconian approach to user management. It assumes guilt by keyword association.
The incident sparked immediate backlash. Critics point out a massive double standard. Anthropic trained its models on scraped public data, books, and code from the open web. Now the company aggressively bans paying users who try to learn from its outputs. Anthropic sells itself as a safe, independent alternative to big tech. Yet it enforces the exact same rigid enterprise lockdowns to protect its commercial moat. Protecting a business model is one thing. Moralizing it while standing on the scraped knowledge of the internet is another.
Key facts
- Opus 5.5 — Model version downgraded to Opus 5 as an automated penalty
- Jiahao Shao — CS PhD researcher whose account was restricted by Anthropic
- RFT and SFT — Prompt topics that trigger automated session kills
Why it matters
Foundation model providers are tightening their terms of service. They want to protect their commercial moats at all costs. Relying on Claude is now a massive operational risk if you build infrastructure or automation tools for AI training. Anthropic will aggressively police your prompts for specific keywords. They will shut off your API access or downgrade your model the moment they suspect you are helping a rival system. You do not own your workflow when you rent access from a paranoid vendor. Your business continuity is at the mercy of their automated filters. One wrong prompt can halt your entire engineering team.
This aggressive policing accelerates the push toward local and open-source models. Builders cannot risk having their research pipelines killed in the middle of a critical job by an opaque moderation filter. Closed models are becoming increasingly hostile to AI engineers. Startups will be forced to migrate their internal tooling to models they actually control. The era of using closed APIs to bootstrap open-source competitors is rapidly ending. Open weights are no longer just a philosophical preference. They are a hard business requirement for anyone building AI developer tools.
For builders
Stop using Claude for model training pipelines
Anthropic is actively scanning prompts for terms related to fine-tuning and model training. If your startup uses Claude to orchestrate training runs, you risk sudden account suspension. Move these critical workloads to open-source alternatives immediately. Startups lose time and money when vendors kill sessions mid-task.
Expect aggressive moat protection from vendors
Closed model providers will not let you use their tools to build competing products. You pay for access, but they completely dictate your use case. Factor sudden bans into your risk models if you build developer tools. The vendor always wins in their own ecosystem.
Audit your reliance on closed model APIs
Your entire research automation pipeline can break without warning if a vendor changes their safety filters. Build fallback mechanisms using local models. This ensures your infrastructure survives arbitrary vendor bans. Founders pay the price when they lack control over their core stack.
My take
Anthropic scraped the open web to build Claude. Now they ban users for automating their own research. This is peak hypocrisy. You cannot claim the moral high ground on AI safety while running a corporate protection racket to crush open-source developers. Stop renting your brain from companies that view your innovation as a threat. Own your infrastructure.
Original reporting: Hacker News. This is my rewrite and opinion.