White Circle raises $11M from the very labs it's built to police

verified

Tuesday, May 12, 2026

The News

White Circle, a platform that helps companies monitor, secure and control AI models in production, has raised $11m in Seed funding, the company announced today; the capital will fund product development and hiring across the US, UK and Europe. The round features personal investments from Romain Huet (OpenAI); Dirk Kingma (formerly OpenAI, now at Anthropic); Guillaume Lample (Mistral); Thomas Wolf (Hugging Face); Olivier Pomel (Datadog); François Chollet (creator of Keras); Mehdi Ghissassi (formerly DeepMind); Paige Bailey (DeepMind); and David Cramer (Sentry) — a list that reads as senior practitioners from the labs that build the models White Circle is designed to police. The product is a single-API control layer that scans AI inputs and outputs in real time against customer-defined policies, detects harmful content, catches hallucinations, blocks prompt-injection attacks, flags model drift and identifies abusive users. Its KillBench study ran more than one million experiments across 15 AI models from OpenAI, Google, Anthropic and xAI and documented that structured-output integrations — the standard for production AI deployments — caused refusal rates to collapse and biases to amplify. Coverage in Fortune and The Next Web.

The Read

$11M is small money; the signal isn't the round, it's the cap table. When the head of developer experience at OpenAI, an Anthropic researcher, Mistral's chief scientist, and Hugging Face's CSO all personally write checks into a company whose pitch is "we sit between your users and the foundation models and catch what they do wrong," they're conceding a thing the labs won't say in press releases: model-side guardrails don't survive contact with production. The KillBench finding is the part B2B buyers should actually internalize — refusal rates collapse the moment you wire a model into a JSON-shaped form, which is how essentially every enterprise deployment looks. If you're running an agent in structured-output mode, you are running it with weaker safety than the demo had.

Counterview

The same pattern produces an arbitrage problem — if the labs improve their own guardrails by 50%, an external control layer becomes redundant table stakes, not a durable business.

Do This Week

Spend two hours this week running your own structured-output regression: take 50 prompts your model currently refuses in chat mode, wrap them in the JSON schema your production app actually uses, and re-run them. Log how many now pass. That delta is the gap White Circle (and Lyrie, and the rest of the control-layer cohort) is selling against — and it's also the gap your CISO will ask you to quantify within the next quarter.

For Product / ML lead