A Paris-based startup founded by a researcher who broke major AI models with a single prompt has raised $11 million to help companies control their own AI systems.
Back
AI safety startup White Circle raises $11 million from industry leaders
(P1) French cybersecurity startup White Circle has raised $11 million in a seed round to expand its AI control platform, a new layer of security for businesses deploying large language models. The funding comes as enterprises grapple with risks beyond the general safety filters built by model providers, such as data leakage, hallucinations, and prompt injection attacks.
(P2) "Jailbreaks are just one part of the problem," Denis Shilov, founder and CEO of White Circle, told Fortune. "In as many ways people can misbehave, models can misbehave too. We’re actually enforcing behavior."
(P3) The investment round saw participation from a roster of AI industry leaders, signaling strong confidence in the company's approach. Backers include OpenAI's Head of Developer Experience Romain Huet, Anthropic's Durk Kingma, Mistral co-founder Guillaume Lample, and Hugging Face co-founder Thomas Wolf. The company will use the funds to hire new talent, accelerate product development, and grow its customer base across the U.S., U.K., and Europe.
(P4) The need for a dedicated control layer becomes more urgent as companies move from simple chatbots to autonomous AI agents that can take actions like accessing files or issuing refunds. White Circle’s platform sits between a company’s users and its AI models, monitoring all inputs and outputs in real time to enforce custom policies, a function model providers themselves have mixed incentives to perfect.
The Paris-based startup was born out of Shilov's own research in late 2024, when he developed a universal jailbreak prompt that could bypass the safety guardrails of every leading AI model. After his findings went viral on X, major AI labs like Anthropic invited him to test their models, convincing him that companies needed more than just better models—they needed better controls.
White Circle’s software is designed to catch a wide range of model misbehavior. It can block a user’s attempt to generate malware, flag when a model hallucinates and promises a customer an unauthorized refund, or stop an AI agent from leaking sensitive financial data. The platform has already processed more than one billion API requests and is in use by the startup Lovable and several fintech and legal companies, though specific revenue figures have not been disclosed.
To bolster its credibility as an independent arbiter, White Circle’s research arm published "KillBench," a study that ran over a million experiments on 15 AI models from vendors including OpenAI, Google, and xAI. The study revealed hidden biases that emerge in high-stakes scenarios, highlighting risks that are not apparent in ordinary use.
This investment is a strong bet on the "picks and shovels" economy growing around the AI sector. The participation of executives from competing AI labs like OpenAI, Anthropic, and Mistral suggests a broad consensus that a third-party enforcement layer is a necessary component of the enterprise AI stack. This points to the emergence of a new, critical market category for investors focused on AI infrastructure and safety.
This article is for informational purposes only and does not constitute investment advice.
[1] AI startup White Circle completes $11 million seed funding round, with participation from OpenAI executives[2] White Circle Raises $11 Million for AI Control Platform[3] White Circle raises $11M to help companies secure and monitor AI model behavior[4] Circle Gives AI Agents USDC Stablecoin Powers Alongside $222M Arc Token Sale[5] Exclusive: White Circle raises $11 million to stop AI models from going rogue in...
