White Circle, a startup specializing in the supervision and security of artificial intelligence models, announces that it has raised 11 million dollars, or approximately 9.35 million euros. The tour brings together several major figures in the global AI ecosystem, including Romain Huet, Dirk Kingma, Guillaume Lample, Thomas Wolf, François Chollet, Olivier Pomel and Paige Bailey.
Founded by Denis Shilov, White Circle develops a platform allowing companies to monitor the behavior of their AI models and autonomous agents in real time. The startup claims to already process more than a billion API requests and counts several large international groups among its clients, including two major global digital banks as well as Lovable.
This lifting comes as companies accelerate the deployment of AI agents in critical environments, whether in customer relations, finance, cybersecurity, human resources or even software automation. The market is gradually moving from the development of models to their operational supervision.
Before White Circle, Denis Shilov made himself known in the AI ecosystem for having published in 2024 a “universal jailbreak” capable of bypassing the security mechanisms of the main generative models on the market. According to the company, this prompt made it possible to obtain responses normally blocked by systems like OpenAI or Anthropic. Safeguards meant to prevent the generation of dangerous, illegal or sensitive content could be overridden with a single instruction.
The post quickly went viral and surpassed 1.4 million views. It also attracted the attention of several American AI laboratories. Denis Shilov was then invited to participate in Anthropic’s bug bounty program before launching White Circle.
White Circle builds an observability and control layer intended to fit between AI models and business applications. Its platform analyzes the inputs and outputs of models in real time in order to detect hallucinations, identify prompt injections, monitor behavioral deviations, block malicious actions or even prevent leaks of sensitive data.
Companies can set their own control policies to determine what is allowed or prohibited. White Circle also offers automatic limitation, blocking or banning mechanisms. The system is designed to work with different AI model providers through a single API. The company claims to support over 150 languages.
In the uses described by White Circle, the platform can prevent an AI agent from executing destructive commands, detect abnormal behavior in financial workflows or identify manipulation attempts aimed at circumventing the internal rules of a model.
The development of the startup illustrates a rapid evolution of the generative AI market. After the race for GPU models and infrastructures, a new segment is emerging around the supervision of AI systems in production. The generalization of low-code tools and “vibe coding” platforms is greatly accelerating the deployment of AI applications in companies. Non-specialized teams can now connect models to databases, CRMs, ERPs or financial tools in a matter of hours.
This democratization, however, creates a new area of risk. Companies must now manage agents capable of interacting with critical systems without always having clear visibility into their actual behaviors. The risks concern the exposure of sensitive data as well as operational deviations, malicious manipulation or the execution of unforeseen actions.
White Circle positions itself precisely on this layer of operational supervision. The parallel with the cloud market is obvious. After the explosion of cloud infrastructures in the 2010s, companies like Datadog or Sentry have established themselves as essential observability layers for monitoring distributed architectures. White Circle is now trying to fill a comparable role for AI systems.
The startup also seeks to strengthen its technical credibility by publishing research on the risks associated with generative models. In 2025, it published “CircleGuardBench”, a benchmark intended to assess the robustness of AI moderation models in real-world conditions. More recently, White Circle presented “KillBench”, a study based on more than a million experiments carried out on fifteen models including OpenAI, Google, Anthropic and xAI.
According to the startup, this work has highlighted behavioral biases linked to nationality, religion, physical appearance and even certain cultural markers. The study also claims that certain structured formats used in enterprise AI integrations greatly reduce the opt-out mechanisms built into the models.
The lifting of White Circle finally comes in a context of progressive tightening of regulatory requirements around artificial intelligence. Companies must now demonstrate their ability to trace the decisions of models, control the actions of agents, document deviations and limit the legal risks linked to automation.
With the emergence of AI agents capable of directly interacting with operational systems, behavioral monitoring of models could become a structural component of enterprise AI architectures. White Circle is betting precisely on this development.
