Cursor's latest AI agent has been exposed for allegedly "shelling" Anthropic's underlying technology—a practice where developers essentially wrap another company's core product with minimal changes. Reverse engineering by developers revealed that Cursor 3.0 runs a local replacement engine that simply swaps out all mentions of "Claude" in system prompts and messages with "Cursor." The installation package directly includes Anthropic's official agent development tools and connects to a Claude 3.7 model that's been fine-tuned specifically for Cursor.
https://twitter.com/jasonkneen/status/2043435856849940818
Official Response to Community Backlash
Facing community skepticism, Cursor team member Michael Truell quickly issued a clarification. He stated this was merely a routine A/B performance test designed to compare the strengths and weaknesses of Claude's official framework versus Cursor's default setup. The test was intentionally discreet, affecting less than 1% of total user traffic.
Michael's explanation appears factual, though the manner of its discovery remains somewhat awkward for the company.