What happened with Figma and Anthropic?
Figma announced a partnership with Anthropic in February 2026 to bring AI-generated work into its design canvas. In April 2026, Anthropic launched Claude Design, a product squarely in Figma's lane. Activist shareholder Findell Capital then pressed Figma to review its governance and the Anthropic relationship over the conflict of interest.
Reported by CNBC and Yahoo Finance. You do not need to assign blame to see the shape of it: a company built on a model provider found that provider moving into its market.
Why this is structural, not a one-off
Because the model providers are incentivized to move up the stack into the applications built on them, and every integration teaches them how. When your supplier's roadmap points at your market, the ordinary act of using them, sending prompts, sharing workflows, exposing what good looks like, is also the act of training your future competitor.
This is the same dynamic playing out across the data economy. The companies selling training data and evaluations to frontier labs are watching those same labs build their replacements. When a market's biggest customer is also building your replacement, your product is not the moat. Your position is, and your position depends on what you keep to yourself.
What does this have to do with AI evaluation?
Your evaluation criteria are the most concentrated form of the thing you should not hand over. They encode what your business considers good, where your current system fails, and what a winning answer looks like in your domain. Give your AI vendor a clean copy, and you have handed it the roadmap to compete with you, distilled.
This is why the answer is not to distrust any particular vendor. It is to change what you expose. An evaluation you own and keep private lets you keep using the best models while withholding the one thing that would help a provider build past you. See why your eval criteria are IP.
How to keep working with vendors without arming them
Use the models. Keep the judgment. Run your evaluation inside your own perimeter, on your data, with criteria the vendor never sees, and swap models freely underneath a loop you own. The vendor sees outputs being tested; it never sees the standard. That is the practical meaning of private, owned evaluation, and it is what PrivateEval builds. See what private AI evaluation is.