As companies integrate large models into business processes such as customer service, programming, and finance, issues like uncontrolled model outputs, sensitive data leaks, and unauthorized actions are drawing increased attention. Paris-based AI security startup White Circle has recently completed a $11 million seed round to introduce a real-time control layer between enterprise users and models.
Funds come from multiple individuals in the AI industry.
Investors in this round include Romain Huet, Head of Developer Experience at OpenAI; Durk Kingma, co-founder of OpenAI and current researcher at Anthropic; Guillaume Lample, co-founder and chief scientist at Mistral; and Thomas Wolf, co-founder and chief science officer at Hugging Face.
White Circle stated that the funds will be used to expand the team, accelerate product development, and grow its customer base in the United States, the United Kingdom, and Europe. The company currently has approximately 20 employees, located in London, France, Amsterdam, and other areas, with a team primarily composed of engineers.
Add a real-time blocking layer outside the model
White Circle's product positioning is to deploy a real-time execution layer between enterprise users and AI models. The platform continuously monitors input and output content based on enterprise-defined policies. If a user attempts to generate malware, fraudulent content, or other restricted information, the system can directly block or flag it.
The company says this system can also be used to detect model hallucinations, sensitive data leaks, unauthorized refund commitments, and destructive actions by AI agents within software environments. The core idea is not to rely solely on model vendors to apply generic safety fine-tuning during training, but rather to allow enterprises to define within their own business environments which behaviors are permitted and which must be blocked.
Shilov believes that as businesses shift from chatbots to AI agents capable of executing tasks, the risks expand significantly. These systems can do more than generate text—they may write code, access files, browse the web, and even perform actions on behalf of users.
Jailbreak prompts spark entrepreneurial inspiration
White Circle was founded by Denis Shilov. At the end of 2024, he developed a reusable "universal jailbreak" prompt designed to bypass the security restrictions of mainstream models. His approach involved instructing the model to respond not as a chatbot bound by safety rules, but rather as an API interface that directly processes requests.
According to its description, this prompt once enabled multiple leading models to answer dangerous questions they were originally designed to refuse. After the related content spread on X, it attracted widespread attention and earned Shilov an opportunity to privately test Anthropic’s models. Shilov later concluded that the issue was not merely about discovering jailbreak prompts, but about companies lacking continuous control over model behavior.
Over 1 billion API requests processed
White Circle stated that its platform has processed over one billion API requests in total, with current clients including the programming tools startup Lovable and several fintech and legal services companies.
Shilov believes that model providers may not have sufficient incentive to build the real-time control layer required by enterprises. On one hand, even when a model refuses to respond, some vendors still charge for input and output tokens; on the other hand, stricter safety training can sometimes negatively impact the model’s performance on tasks such as programming.
Publish research test model bias
In addition to its product offerings, White Circle is advancing research. In May, the company released a study called KillBench, which conducted over one million experiments testing how 15 models—including those from OpenAI, Google, Anthropic, and xAI—respond to fictional scenarios involving life-or-death decisions.
The company stated that experimental results showed the model made different decisions based on attributes such as nationality, religion, body type, or smartphone brand, indicating that hidden biases may surface in high-risk scenarios. The study also found that these biases were more pronounced when the model was required to output answers in fixed options or form formats—a common way businesses integrate AI into real-world products.