All analysis was generated autonomously, without human review. Scores are analytical opinions drawn from the cited public sources, without hands-on testing. They are not audits, certifications, investment reports, purchasing advice, or evaluations of quality.
White Circle sells companies an AI control platform for the models and agents in their products. It checks inputs, outputs and agent actions in real time against policies each customer writes in plain language. Its named customer is the coding startup Lovable, whose CISO says the product protects millions of AI interactions a day. The company says it has a multi-million-dollar run rate, with digital banks and fintech and legal companies as customers. Leaders at OpenAI, Anthropic and DeepMind backed its $11 million seed round. Fortune reports a team of 20, almost all engineers. White Circle bets that model providers will not solve AI safety through training alone. Buyers could drop White Circle if those providers or monitoring vendors such as Datadog add per-customer policy checks.
| Description | White Circle sells an AI control platform that tests AI applications for failure modes, enforces custom policies on inputs, outputs, and agent tool calls in real time, and analyzes how users and models behave over time. | [f1] |
|---|---|---|
| HQ | Paris, France | [f2] |
| Funding | $11M total | [f3] |
| Latest funding | Seed | [f3] |
| Deployment | SaaS | [f4] |
| Compliance | HIPAA, SOC 2 Type 2 | [f4] |
| Product | What it does |
|---|---|
| White Circle | An API and platform that tests AI applications for failure modes, enforces plain-language policies on inputs, outputs, and tool calls in real time, and tracks user and model behavior. |
AI Defense Matrix
| Govern | Identify | Protect | Detect | Respond | Recover | |
|---|---|---|---|---|---|---|
| AI-Workload Platforms Inference servers, training platforms, vector DB platforms, and the model-loading supply chain. | ||||||
| AI Coding and Orchestration Tools AI coding tools and agentic orchestration tools on user devices, plus their plugins, skills, hooks, system prompts, scaffolding, harnesses, configuration settings, and MCP clients. | ||||||
| AI-Generated Code Code produced by AI tools, AI-assisted reviews, AI-generated infrastructure-as-code and tests, and vibe-coded apps that bypass CI/CD. | ||||||
| AI Traffic Paths to external and self-hosted AI services, MCP tool channels, agent-to-agent calls, model-registry downloads, and egress to unapproved AI services. AI gateways, LLM routers, and MCP gateways steer traffic along those paths. The steering decisions and their inputs belong here too: routing policies, MCP server registries, and agent naming and discovery services. | ||||||
| AI Model Model weights, fine-tuning checkpoints, model cards, registries, an AI bill of materials (AIBOM), and the third-party LLMs your enterprise consumes. | ||||||
| Training Data Datasets used for training, fine-tuning, and continued learning. | ||||||
| Runtime AI Data User prompts, inference inputs, RAG content, vector DB content, persistent agent memory, and interaction history. | ||||||
| AI Agent Identities AI agents as non-human principals, plus credentials, keys, permission scopes, service accounts, and delegation chains across agents and tools. |
White Circle enforces customer-written policies on the inputs, outputs, and tool calls of AI applications and tests AI systems for failure modes before users meet them. These capabilities are mapped to the AI Defense Matrix. [f5]
How well the company can compete in its security market, scored across eight dimensions against public evidence.
| Dimension | Score | Rationale |
|---|---|---|
| Problem Clarity How precisely the company defines its problem, with evidence the problem exists at the scale claimed. | 3/5 | White Circle names a clear buyer, companies running AI models and agents inside their own products, and concrete failure modes such as prompt injection, hallucination and abusive users (s11, s12). The quantified evidence of the problem is KillBench, the company’s own bias study across 15 frontier models (s5), so the pain is clearly stated and credible but not independently quantified. [s11, s12, s5] |
| Capability Depth How specific the technical capabilities are, with evidence beyond marketing claims such as docs and third-party validation. | 3/5 | Public API documentation covers policies, conditions, environments and user-level strikes, and the Lovable case study details how events flow through the API (s6, s2). CircleGuardBench ranks White Circle’s own moderation models against other guard models (s4). White Circle published both the comparison and the Lovable case study. Neither is a third-party evaluation of the product’s detection. [s6, s2, s4] |
| Market Timing Whether the market is ready for this product, with evidence that buyers are actively seeking solutions. | 3/5 | The enabler is the move from chatbots to autonomous AI agents that write code and take actions, which Denis Shilov describes to Fortune after his late-2024 universal jailbreak (s11). The buyer-side demand signals are of one kind, the company’s own customer statements and its Lovable case study (s10, s2), short of multiple independent buyer-side signal kinds. [s11, s2, s10] |
| Team Credibility Demonstrated domain expertise with public signals such as prior exits, publications, and industry recognition. | 3/5 | Denis Shilov’s 2024 universal jailbreak drew independent press coverage and an invitation from Anthropic to test its models privately (s9, s11), and the company has published a run of research including CircleGuardBench and KillBench (s4, s5). The research is self-published and no prior in-domain product build or exit for the founder is documented, so this is verifiable relevant experience short of the plural, independently evidenced record the next rung needs. [s9, s11, s4, s5] |
| GTM Proof Evidence of actual traction (customers, revenue signals, partnerships) beyond stated intentions. | 3/5 | Lovable is a named customer, with a July 2026 case study and a quote from its CISO on White Circle’s site (s2, s1). The billion-request figure, the two large digital banks, the fintech and legal customers and the multi-million-dollar run rate are the company’s own statements (s10, s11, s15). TechFundingNews and Fortune relay them and White Circle’s job listings post them, so the score reflects one named reference without independent corroboration of scale. [s2, s1, s10, s11, s15] |
| Funding Efficiency Whether funding matches go-to-market ambition, with signs of capital-efficient growth. | 3/5 | SiliconANGLE and SecurityWeek report an $11 million seed round (s9, s12), and the company ships a documented product that Lovable uses (s6, s2). The multi-million-dollar run rate appears in White Circle’s own job listing (s15), so output per dollar is visible but unconfirmed, the default for a funded private startup. [s9, s12, s6, s2, s15] |
| Category Clarity Whether the company creates or fits a recognizable category that buyers can quickly place in their stack. | 3/5 | Outlets place White Circle under different labels (s9, s12, s13, s10). SiliconANGLE calls it an AI guardrail and monitoring startup, SecurityWeek an AI control platform and Tech.eu an AI governance platform, while TechFundingNews puts it in AI observability and governance alongside Arthur AI and WhyLabs. Buyers can place it, and the shifting labels mark a category still forming. [s9, s12, s13, s10] |
| Incumbent Defensibility How vulnerable the core value proposition is to absorption as a feature by a platform vendor. | 3/5 | Inline integration, customer-written policies and backtesting against a customer’s own historical traffic give White Circle some friction against absorption (s2, s6). TechFundingNews notes that Datadog and Scale AI have expanded into AI monitoring (s10), and no cited source documents a data or procurement position that bundling could not eventually match. [s2, s6, s10] |
White Circle serves companies that put AI models and agents into their own products and need to control what those systems do once people use them. Fortune describes the product as software that checks the inputs and outputs passing between a company’s users and its AI models in real time against company-specific policies.
The failure modes are concrete. SecurityWeek lists hallucinations, harmful content, prompt injection attacks, abuse and model drift among the risks the platform monitors. White Circle’s Lovable case study describes users trying to generate phishing websites and malware through Lovable’s coding agents, which generate code, run commands and take actions for users in real time.
White Circle’s own research quantifies part of the problem. Its KillBench study tested 15 frontier models from nine providers and reports statistically significant biases in every one. The company publishes that work alongside its stated position, reported by Fortune, that AI safety will not be solved entirely at the model-training stage. [s11, s12, s2, s5]
White Circle sells one platform built around four functions its homepage names. They are testing AI applications for failure modes, enforcing behavior across inputs, outputs and tools, analyzing how users and models behave over time, and tuning prompts and models from live signals.
Enforcement is the security core. Customers write policies as plain-language rules, and White Circle’s product routes each event to the right policies through conditions and separates development, staging and production environments. White Circle’s comparison page says the product inspects tool arguments, tool outputs, agent traces and fetched content, and its features page describes a strike system that tracks violations per user and suspends repeat offenders.
Detection runs on White Circle’s own models. SiliconANGLE reports the company employs its own specialized AI models. The Lovable case study describes proprietary self-adjusting classifiers that let a customer test a new policy against historical traffic before pushing it to production. The homepage lists support for more than 150 languages. [s1, s6, s20, s21, s9, s2]
Press places White Circle in the AI observability and governance market. TechFundingNews writes that it operates alongside Arthur AI and WhyLabs in that market, while Datadog and Scale AI have expanded their own AI monitoring capabilities.
White Circle names a rival itself. It publishes a comparison page titled “Why teams choose White Circle over Lakera” and a second page contrasting its custom policies with the fixed detector sets of typical guardrail vendors. Both pages are the company’s own positioning.
The product works independently of any one model provider. The Lovable case study says White Circle is model-agnostic and that Lovable can switch providers or upgrade models without rebuilding its supervision layer. Denis Shilov told Fortune that model labs do some safety tuning but that it is very general. White Circle sells enforcement of each customer’s own policies on top of whichever model the customer runs. [s10, s19, s20, s2, s11]
Lovable is White Circle’s named customer. White Circle published a Lovable case study in July 2026, and its homepage quotes Lovable’s CISO, Igor Andriushchenko, saying White Circle protects millions of AI interactions every day.
White Circle’s wider traction figures are its own statements. TechFundingNews relays the company’s statement that it has processed more than one billion API requests and serves two major digital banks, and Fortune relays that it is used by several fintech and legal companies. White Circle’s job listings describe a multi-million-dollar run rate and signed customers including Lovable and multiple neobanks.
White Circle is building out direct sales alongside the developer API. The homepage advertises a five-minute API integration, and the company is recruiting its first dedicated US sales hire and a founding GTM engineer. [s2, s1, s10, s11, s15, s8]
Founder and CEO Denis Shilov came to White Circle from AI jailbreak research. SiliconANGLE reports he went viral in 2024 after showing he could jailbreak numerous proprietary AI models with a single prompt, and Fortune reports the post brought an invitation from Anthropic to test its models privately.
The company has published a series of research projects. It released CircleGuardBench, a benchmark for AI moderation models, and KillBench, a bias study across 15 frontier models. It has also published a fine-tuned coding model dated September 2026 and Halo, an open-source distributed training project. That work appears on White Circle’s own research pages.
Fortune reports a team of 20, almost all engineers, spread across London, France, Amsterdam and elsewhere in Europe. The seed round’s named backers include OpenAI’s head of developer experience Romain Huet, OpenAI cofounder Durk Kingma, Mistral cofounder Guillaume Lample, Hugging Face cofounder Thomas Wolf and Datadog chief executive Olivier Pomel. [s9, s11, s4, s5, s3]
White Circle’s homepage lists SOC 2 Type II and HIPAA, and TechFundingNews reports the company is SOC 2 Type I and II certified and HIPAA compliant. Its terms of service identify the contracting entity as White Circle, PBC, a Delaware public benefit corporation.
Auditability is built into the product. The Lovable case study lists audit logs for compliance and incident review and a human review queue for edge cases. The homepage says customers can see what happened, why it happened and which policies were triggered.
The platform runs as a hosted API with separate EU and US regions, according to its documentation. White Circle’s comparison page against Lakera publishes accuracy figures and offers the test set and reproduction steps on request, so a buyer can check the claims before relying on them. [s1, s10, s18, s2, s6, s19]
| Company | Relationship | Note | Compare |
|---|---|---|---|
| Lakera | competes with | White Circle publishes a comparison page against it. | |
| Arthur AI | competes with | TechFundingNews places it in the same AI observability and governance market. | |
| WhyLabs | competes with | TechFundingNews places it in the same AI observability and governance market. | |
| Datadog | competes with | TechFundingNews names it among monitoring companies that expanded AI monitoring capabilities. | N/AWe scored these companies at different scopes, so the totals measure different things. |
| Prompt Security | competes with | Competes for the same buyer of runtime guardrails for AI applications. |
Add analyzed competitors to compare them side by side with White Circle.
A closer look at the company's product strategy, measuring how defensible it is against market forces and examining the eight areas behind it.
reinforce or reposition
White Circle runs in the request flow of the AI applications it protects. Lovable sends every session event to its API and defines policies tuned per feature, so a customer that leaves has to re-route those calls and move its policies to another vendor. Its classifiers adapt to each customer through labelled feedback. Its SOC 2 and HIPAA attestations are commercial credentials a funded rival can also obtain. By the company’s own account, its customers include two major digital banks, which puts regulated financial institutions in its buyer base.
| Dimension | Score | Rationale |
|---|---|---|
| Value Delivery Does the product sell software as the product, or judgment, trust, or accountability with software as the delivery mechanism. | 1/3 | Customers write the policies and operate White Circle through its API, dashboard and review tools (s6, s2). White Circle sells software the customer’s team runs and owns the outcomes of, which is the software-product level. |
| Switching Cost How expensive leaving is for a customer: data portability, integrations, learned workflows, network effects, regulatory data residency. | 2/3 | Lovable sends every session event to White Circle’s API and runs custom policies tuned per feature and use case, so leaving means re-routing those calls and moving the custom policies to another vendor (s2). White Circle’s documentation describes no policy export and does not size the migration work. |
| Compliance Moat Whether certifications, liability acceptance, or audit trails block an easy replacement. | 1/3 | White Circle’s homepage lists SOC 2 Type II and HIPAA, and TechFundingNews reports SOC 2 Type I and II certification and HIPAA compliance (s1, s10). A funded competitor can obtain these commercial attestations through ordinary enterprise preparation, so they are entry cost that a replacement vendor can also meet. |
| Problem Complexity Whether the product requires ML, optimization, real-time systems, or years of specialized expertise. | 3/3 | White Circle runs its own proprietary classifiers inline on every event, cites 50 to 100 millisecond end-to-end latency, and supports more than 150 languages (s2, s20, s1). Its research team also trains and evaluates models (s3, s4), which places the product in machine-learning and real-time systems work that takes specialized expertise. |
| Buyer Profile Whether buyers are SMB operators, mid-market IT teams, or regulated enterprises and governments with procurement gates. | 3/3 | White Circle states that two major digital banks use its platform (s10), and its job listing cites multiple neobanks among signed customers (s15). That places regulated financial institutions in its buyer class alongside the software platform Lovable (s2). The banks are unnamed and the claim is the company’s own, so the rung depends on White Circle’s account of those customers. |
| Layer Whether the product is an end-user application, a platform with application features, or infrastructure other applications depend on. | 3/3 | White Circle runs inside its customers’ applications. Lovable sends every new event in a session to White Circle’s API endpoint, which evaluates and enforces policies in real time (s2). White Circle’s comparison page says inline deployment makes latency and uptime part of the customer’s stack (s20). The dashboards, testing and review tools sit on top of that enforcement role (s1, s2). An application can process an event with no policies attached (s6), so the dependence is on enforcement, not on the application running at all. |
| Proprietary Data, Content, or IP Whether the product accumulates datasets, content licenses, or IP that a rival cannot recreate from scratch. | 1/3 | White Circle runs its own classifiers (s2), and SiliconANGLE relays the company’s account that its models improve through labelled user feedback and adapt to each use case (s9). That describes a technique and per-customer tuning, and no cited source names a dataset the vendor retains across customers. White Circle publishes code and data for its KillBench bias study (s5). That dataset is public, and no retained, non-public asset is documented. |
White Circle sells to companies that run AI models and agents inside products their own customers use. Its named customer, Lovable, runs coding agents that generate code, run commands and take actions for users. White Circle says it also serves two major digital banks and several fintech and legal companies.
The homepage lists government, healthcare, education and research among the industries it addresses. Those are target segments on the company’s own page, while the customers it describes are software platforms, financial services firms and legal companies.
White Circle’s staff and sales hiring span Europe and the U.S. Fortune reports staff across London, France and Amsterdam, and White Circle’s job listings add New York, San Francisco and London sales roles, matching its stated plan to grow customers across the U.S., U.K. and Europe.
White Circle’s product combines testing, enforcement, analytics and tuning in one platform. The homepage frames these as finding failure modes before users do, enforcing behavior across inputs, outputs and tools, understanding user and model behavior over time, and tuning prompts and models from live signals.
The AI work is in the classifiers. White Circle runs its own proprietary self-adjusting classifiers, according to the Lovable case study, so a customer can add a policy and test it against historical traffic without training a dedicated model for it. SiliconANGLE relays the company’s account that its models improve through labelled user feedback and become better adapted to each use case. SecurityWeek reports the platform can block AI agents from performing malicious actions and flag and ban abusive users.
White Circle’s research covers more than guardrails. Its CircleGuardBench benchmark measures how well guard models detect harmful content, resist jailbreaks, avoid false positives and respond quickly. It has also released a fine-tuned coding model dated September 2026 and Halo, an open-source distributed training project.
White Circle’s go-to-market starts with a developer API. The homepage advertises a five-minute integration, and the documentation walks a developer from a first policy to a first API call.
White Circle is building out direct enterprise selling. Its founding account executive listing describes the role as the company’s first dedicated US sales hire and says the sales motion is not fully figured out. The same listing says the company is at a multi-million-dollar run rate with signed customers including Lovable and multiple neobanks. The careers page also lists a founding events lead and a founding GTM engineer.
White Circle features its backers in its own marketing. The seed round that followed the 2024 jailbreak SiliconANGLE and Fortune describe has named backers including senior people from OpenAI, Anthropic, Mistral, Hugging Face and Datadog.
White Circle keeps its prices inside the product. Its whitecircle.com/pricing address returns a not-found page. Its terms of service say pricing plans are available to workspace administrators in the account settings.
The terms describe usage-based billing. Customers pay fees based on product tier and usage, can buy prepaid credits that deplete per API request, and pay a per-request overage rate once credits run out. The API documentation includes a Usage & Cost endpoint for tracking that consumption.
White Circle delivers its platform as a hosted API. The documentation describes separate EU and US platforms, with API keys prefixed by region. The Lovable case study describes every session event going to a single API endpoint, where White Circle evaluates and enforces policies in real time.
Because White Circle runs inline, its speed and availability become part of the customer’s product. White Circle’s own comparison page states the trade-off plainly, saying inline deployment makes latency and uptime part of the customer’s stack, and it cites 50 to 100 milliseconds end to end.
Customers operate the product through its tools. The Lovable case study lists a real-time dashboard, alerts, audit logs, policy testing before rollout and a human review queue for edge cases.
White Circle presents commercial compliance credentials. Its homepage lists SOC 2 Type II and HIPAA, and TechFundingNews reports the company is SOC 2 Type I and II certified and HIPAA compliant.
The product records what it decides. The homepage says customers can see what happened, why it happened and which policies were triggered, and the Lovable case study lists audit logs for compliance and incident review.
Accuracy claims are the company’s own. White Circle’s comparison page against Lakera publishes F1 accuracy figures and offers the test set and reproduction steps on request. Its terms of service name White Circle, PBC, a Delaware public benefit corporation, as the contracting entity.
White Circle positions itself as a control layer that works across model providers. The Lovable case study says it is model-agnostic and that Lovable can switch providers or upgrade models without rebuilding its supervision layer.
Its integrations are documented at the API level. The documentation lists a Slack integration, webhooks and an Events API that checks messages, tool calls, agent state and artifacts.
White Circle publishes much of its research openly. It hosts CircleGuardBench on Hugging Face, releases code and data for its KillBench study of AI model biases, and shares its training and fine-tuning work on its research pages.
Denis Shilov founded White Circle and runs it as CEO. SiliconANGLE reports he went viral in 2024 after jailbreaking numerous proprietary AI models with a single prompt, and Fortune reports the post brought an invitation from Anthropic to test its models privately.
The team is small and engineering-heavy. Fortune reports 20 people, almost all engineers, across London, France, Amsterdam and elsewhere in Europe, and the careers page lists research, ML engineering and data-labeling roles in Paris and London.
The company is expanding its commercial team after a seed round Tech.eu reports came from leaders at OpenAI, Anthropic, DeepMind, Hugging Face and Mistral. It is recruiting its first dedicated US sales hire and founding events and GTM engineering roles in New York, San Francisco and London, alongside finance and talent leads.
| Id | Source | Tier | Accessed |
|---|---|---|---|
| f1 | White Circle homepage | official | 2026-09-23 |
| f2 | Fortune: White Circle raises $11 million to stop AI models from going rogue in the workplace | press | 2026-09-23 |
| f3 | SiliconANGLE: White Circle raises $11M to help companies secure and monitor AI model behavior | press | 2026-09-23 |
| f4 | AI Defense Matrix Catalog entry | other | 2026-06-14 |
| f5 | AI Defense Matrix Catalog mapping | other | 2026-09-23 |
This site is an experimental research aid created by Zeltser Security Corp. All its data gathering and analysis was performed autonomously without human review, and it can contain errors of fact, interpretation, and judgment that a human reviewer might catch.
The analyses are statements of opinion, not statements of fact. Machine analysis produced the scores, summaries, and matrix placements by weighing the public sources each page cites, and reasonable people can weigh the same sources differently. Where a page states a fact, it cites the public source and the date it was checked, and the statement is only as accurate as that source. Unless a profile expressly says otherwise, the analysis involves no hands-on testing and no independent validation of any company's products or services.
Nothing here is professional, security, legal, financial, investment, or purchasing advice, and nothing here is a recommendation to invest in, do business with, or avoid any company. Inclusion of a company is not an endorsement, and absence of a company is not a judgment about it. Reading this site creates no advisory or client relationship. Verify any detail you plan to act on against the vendor's current materials.
The content is provided "as is" and "as available," with all warranties disclaimed, express or implied, including merchantability, fitness for a particular purpose, accuracy, and non-infringement. No entry is warranted to be complete, current, or correct. Companies change, vendors update their claims, sources can be wrong, and automated analysis can misread them.
To the fullest extent permitted by law, the operator, Zeltser Security Corp, is not liable for any damages that arise from using this site or relying on its content, including direct, indirect, incidental, special, and consequential damages and lost profits, even if advised that such damages were possible. If you are dissatisfied with the site or disagree with these terms, your remedy is to stop using it.
Entries link to vendor pages, press coverage, and other external sites that Zeltser Security Corp does not control and is not responsible for. A link is not an affiliation with the destination or an endorsement of it. Product and company names and trademarks are the property of their owners, used here nominatively to identify the companies described. Short quotations from cited sources appear for identification and commentary.
Use, quotation, automated retrieval, and redistribution of the content are governed by the Terms of Use at cybercompanyprofiles.com/terms, which permit personal and internal business use with attribution and prohibit republication and resale.