Whitecircle
13 known investors
White Circle provides a real-time control layer for AI systems, letting teams test for failure modes, enforce behavior policies across inputs, outputs, and tools, analyze usage, and improve models. It targets enterprises shipping AI in production, offering guardrails around safety, security, compliance, and reliability.
Also known as White Circle · White Circle AI · White Circle, PBC
Founders & leadership
Investors · 13
Also in the syndicate · 12
Company profile
researched Aug 2026White Circle operates an AI control and safety platform that lets companies test, protect, observe and improve AI systems while they run in production. Its proprietary models inspect model inputs and outputs in real time against customer-defined policies, detecting harmful content, hallucinations, prompt injection attempts, model drift, sensitive-data leakage and abusive or malicious users. Enforcement actions can block, rewrite, escalate or terminate risky behavior, trigger alerts, require human confirmation for agent actions, and apply rate limits or bans, with policies varied by user location or jurisdiction and an audit view showing which policies were triggered and why.
Beyond guardrails, the platform provides analytics and optimization layers: custom metrics, risk scoring, error-rate and anomaly analysis, user-satisfaction scoring, prompt engineering, dynamic model routing, memory management and context enrichment. Functionality is delivered through a single API and SDK plus a SaaS platform hosted at regional endpoints (eu.whitecircle.com, us.whitecircle.com and whitecircle.com). The company states 150+ supported languages, unlimited custom policies, a five-minute integration time and 99.99% API uptime, and is reported to be SOC 2 Type I and Type II certified and HIPAA compliant.
White Circle also publishes research on model behavior, including CircleGuardBench, a benchmark for AI moderation and guard model performance released in May 2025, and KillBench, a study of decision-making biases across 15 frontier models. Additional work on distributed training methods, custom GPU kernels and offline reinforcement learning is listed as forthcoming. The corporate entity is White Circle, PBC, a Delaware public benefit corporation.
Founding story
The company was founded by engineer Denis Shilov, who in 2024 published a universal jailbreak prompt that bypassed the safety filters of major AI models, including ChatGPT and Claude, to elicit content those models were built to refuse. The post reached 1.4 million views and attracted attention from Anthropic, OpenAI and Hugging Face; Shilov was invited into Anthropic's bug bounty program and subsequently built the White Circle platform in response to demand from AI providers.
Business model
B2B SaaS. Customers access the platform, API and SDK under a subscription/terms-of-service agreement, with usage consumed through prepaid credits — units of value drawn down per API request — alongside separately negotiated enterprise agreements. Direct competitors are contractually barred from using the service without written consent.
Usage-based access to the API and platform via prepaid credits consumed per API request, alongside subscription and separately negotiated enterprise agreements for the SaaS platform, API, SDK and documentation.
Traction
Reported to have processed more than 1 billion API requests, with customers said to include Lovable and two of the world's largest digital banks; a customer CISO is quoted describing protection of millions of AI interactions daily. The CircleGuardBench article on Hugging Face and several public datasets have drawn community engagement. Headcount is reported at about 30 employees, with 21 open roles listed in August 2026 across Paris, London, New York, San Francisco and remote.
Latest developments
Announced an $11 million seed round on 12 May 2026, with proceeds allocated to product development and hiring across the US, UK and Europe. The company published the KillBench study of decision-making biases across 15 frontier models, listed forthcoming work on communication-efficient distributed training (Hat-Muon), custom kernels for Blackwell B300 GPUs and offline GRPO at scale, and was recruiting for 21 roles as of August 2026, including AI red team, ML infrastructure, DevOps, finance and go-to-market positions.
▸Full profile — market position, technology, go-to-market, geography, history, risks & controversies
Market position
Positioned in the AI security, governance and guardrail segment, described by one profile as competing with vendors such as Arthur AI and WhyLabs, and grouped alongside Pillar Security, Prompt Security, Reken AI and Adaptive Security. The company describes itself as an all-in-one enterprise-grade platform covering testing, protection, observability and improvement rather than a point tool.
The company presents a single API covering the full AI lifecycle — pre-deployment testing, real-time input and output enforcement, analytics and model optimization — rather than separate tooling for each stage, together with customer-defined and jurisdiction-specific policies, automated enforcement actions, auditability of policy decisions, coverage of more than 150 languages, and SOC 2 and HIPAA attestations. Its published empirical research (CircleGuardBench, KillBench) and a founder known for demonstrating cross-model jailbreaks are cited by an investor as evidence of technical credibility.
Technology
The company builds proprietary guard models that evaluate AI inputs and outputs in real time against customer-authored policies, identifying unsafe inputs, jailbreaks, PII, hallucinations, prompt injection, tool abuse, token drain and compliance violations, and verifying output quality, reliability and instruction following. Models improve over time by learning from labelled user feedback and are said to work across more than 150 languages. Surrounding capabilities include dynamic model routing, prompt optimization, context enrichment, risk scoring and anomaly detection, all exposed through one API. Research output includes the CircleGuardBench moderation benchmark and the KillBench bias study, plus published datasets and Spaces on Hugging Face.
Go-to-market
Self-serve sign-up and demo requests through the website, supported by a sales motion (contact sales, founding account executive and partnerships roles) and developer distribution via API, SDK and documentation. Open-source research artifacts, benchmarks and datasets published on Hugging Face and GitHub serve as a technical audience channel, and a trust center with SOC 2 and HIPAA attestations supports enterprise procurement.
Enterprises and product teams deploying AI in production, including regulated and safety-sensitive sectors. The site names government, healthcare, education and research as served industries, and coverage cites fintech and digital banking use cases as well as creative tools and travel.
Geography
Headquartered in Paris, France, with the operating entity incorporated in Delaware as a public benefit corporation. Job postings span Paris, London and US locations (New York and San Francisco) plus remote roles, and seed proceeds are earmarked for team expansion across the US, UK and Europe. The platform is offered via separate EU and US regional endpoints, and policies can be adapted by jurisdiction.
History
Founded in 2025 by Denis Shilov following his viral 2024 jailbreak demonstration. In May 2025 the company published the CircleGuardBench benchmark for AI moderation models, followed by additional public datasets and the KillBench bias study. In May 2026 it announced an $11 million seed round backed by angels from across the AI ecosystem, with the capital directed at product development, team growth in the US, UK and Europe, and customer acquisition. Terms of service for the platform take effect from July 9, 2026.
Risks & controversies
Reported research findings underline risks in the market the company addresses: the KillBench study found that all 15 tested models exhibited preferences tied to nationality, religion, body type and phone brand when making decisions about human lives, and that structured output — the standard production integration method — repeatedly caused refusal rates to collapse and biases to amplify. The founder's public profile derives from disclosing a universal jailbreak affecting major AI models. Sources also show inconsistent basic data on the company, including differing website domains (whitecircle.com versus whitecircle.ai), locations (France versus Delaware) and headcount (1-10 versus 30).
Compiled by commissioned research from 8 cited public sources — announcements, filings, and press listed under research sources below.
Key figures
latest reportedCompany-reported or press-reported figures, each dated to when it was claimed — not independently audited.
Timeline · 4
launches, deals, and filingsWhite Circle announced an $11 million seed round to accelerate product development, expand its team across the US, UK and Europe, and grow its customer base. Investors named include angels from OpenAI, Anthropic, Mistral, Hugging Face, DeepMind, Datadog and Sentry; F4 lists Hummingbird VC as leading the round alongside Factorial, Nomads and Abstract Ventures.
$11M source ↗
White Circle is reported as SOC 2 Type I and Type II certified and HIPAA compliant.
The company published KillBench, a study running more than one million experiments across 15 AI models from providers including OpenAI, Google, Anthropic and xAI, reporting preferences tied to nationality, religion, body type and phone brand, and finding that structured output collapsed refusal rates and amplified biases.
White Circle published CircleGuardBench, a benchmark measuring how LLM guard/moderation models perform on safety and accuracy under real-world conditions, released with an accompanying Hugging Face Space and article.
Dated company events from announcements, filings, and press; legal rows summarize public dockets and regulator releases.
▸Research sources · 8
primary sources listed
- Whitecirclewhitecircle.com · web
8 public sources were cited for this profile; the first-party ones are listed here.
Frequently asked questions
- What does Whitecircle do?
- White Circle is an AI control platform that tests, protects, monitors and optimizes AI systems in production via a single API.
- Who founded Whitecircle?
- Whitecircle was founded by Denis Shilov.
- Who are Whitecircle's investors?
- Whitecircle's investors include Moonfire Ventures.