An AI assistant for business is a Claude-powered pipeline that handles one process - leads, documents, invoices, reports - under your account and your policy. Anthropic is now terminating accounts whose work merely touches training larger models, even when Claude is only used to submit jobs and analyze the results.
A researcher writing on X says his rollouts came from a strong open-source model and Claude only helped with job submission and analysis, yet he was flagged for training a competitor. The account got Opus 5.5 downgraded to Opus 5, sessions killed outright, and anything touching RFT, SFT or training larger models now gets blocked.
What this means for small business automation
If your Claude agents fine-tune, distill or even orchestrate other models, you are one classifier false positive away from a production outage. For a small business running an automation on Claude, two things follow: keep business pipelines (document OCR, lead routing, support) on a separate account from any ML research work, and wire in a fallback provider during the prototype build. Don't share one API key between a chatbot in production and a researcher's SFT notebook.
Why it matters: Vendor policy is now part of your architecture. If Anthropic can kill a session for touching the word training, your automation needs a second lane ready before the ban, not after.
Source: Hacker News