Enterprises are handing AI vendors something more dangerous than access: their data. On July 13, a security researcher reverse-engineered xAI's Grok Build binary and found it uploaded a developer's entire codebase to xAI's cloud during a session with zero AI calls. xAI disabled the behavior after the finding went public. The same week, Satya Nadella wrote that enterprises 'pay for intelligence twice, once with money, and again with proprietary knowledge.' Alex Karp told CNBC's Squawk Box that frontier labs are 'stealing the weights and alpha of my business.' Two major Microsoft and Palantir executives, both partners and competitors to the frontier labs, issued the same warning in the same week.
The mechanism is called a trajectory: the data produced each time a user queries an AI through a harness like Cursor or Claude. In SaaS, customer data sat in databases the vendor could not touch. Trajectories are different. They can be fed back into model training, converting enterprise data into vendor intellectual property. The data pipeline industry built on this reality now generates roughly $10 billion in revenue. The original piece maps exactly how the inference loop and training loop connect through the harness layer, and that diagram alone is worth the read.
The next contractual standard will be zero data retention, not anonymization. Anonymization techniques are not yet strong enough to guarantee protection. The argument the original makes is structural: the SaaS era spent 20 years proving vendors could be trusted with data storage. The AI era will have to make the same guarantees from scratch, and the harness is where that fight will be won or lost.
[READ ORIGINAL →]