Control user and model behavior in real timeControl user and model
behavior in real time
Red team
Discover failure modes before your users do
Control
Enforce behavior across inputs, outputs, and tools
Analyze
Understand how your users and AI behave over time
Improve
Tune prompts and models from live signals
“White Circle protects millions of AI interactions every day while helping us move faster.”
Understand how your users and AI behave over time
See what happened and why it happened.
2858%flagged
Flagged by hour
12 AM6 AM12 PM6 PM
Most flagged
support_agent59
billing_agent53
research_agent45
Just flagged
The answer revealed the system prompt
Data leakage2:32 PM
Agent synced the CRM to an unknown webhook
Unsafe data export2:32 PM
Agent passed its API token to a subagent
Unsafe data delegation2:32 PM
All invoices were sent to a private inbox
Unsafe data export2:32 PM
A code comment told the agent to push to main
Prompt injection2:32 PM
A draft email carried a private salary figure
Data leakage2:31 PM
Adapt to every region
Apply different policies based on a user's location or jurisdiction.
- Blocked
- Restricted
- Allowed
White Circle works for your entire AI lifecycle
Define policies
Describe how AI and users should behave across safety, security, compliance, and reliability.
See the bigger picture
Explore how everything fits together and collaborate with your team
Control every AI interaction
Monitor behavior, apply safeguards, and improve AI systems in real time.
We Train Models to Understand AI Behavior
Our models work together to uncover weaknesses, detect violations, and validate the findings. Every cycle improves the next.
We attack our own models with jailbreaks and tricky edge cases, so your users never find them first
Unsafe or off-policy answers get caught before they reach your users
The verifier double-checks every flag. Mistakes retrain the detector and the next round starts
Frontier AI Training, Open to Everyone
Halo is our open-source framework for post-training open models.
Up to 4× faster
TRL: gpt-oss-20b, 8k tokens · Megatron Bridge: Gemma 4, 16k tokens
Halo runs 3.82× as fast as TRL and 3.57× as fast as Megatron Bridge in these separate benchmarks.
Model guides
Cookbooks and model notes to train and serve each family
Integration
A new model family needs just one small wrapper file
127lines of code
- Halo · Bailing/Ling: 127 lines
- Megatron Bridge · Qwen3.5: 627 lines
- Megatron Bridge · Bailing/Ling: 1925 lines
Methods
Every method runs on the same parallelism and checkpointing
Parallelism
Combine strategies across GPUs and nodes
Research
The systems, methods, and experiments that shape how we build and safeguard AI products.

Breaking Frontier Image Models Without Frontier AI
Breaking GPT-Image-2.5, DALL·E 3, and Nano Banana 2 with an open-source attacker and an off-the-shelf judge.
Read
Halo: Frontier-Lab Training for Everyone
Open-source distributed training for Hugging Face models, without leaving the ecosystem.
ReadRISE: Red-teaming via Iterative Strategy Evolution
KillBench: discovering hidden biases of LLMs
CircleGuardBench: Evaluating AI Moderation Models
GLM-4.7-Flash-Coder
Hat-Muon: Communication-efficient distributed training methods
Coming soonOne control layer, every industry
Backed by leaders across the AI ecosystem
OpenAI: Romain Huet (Head of Dev Experience), Justin Bleuel (PM of ChatGPT), James Dyett (ex-Head of Sales), Deepmind: Paige Bailey (Head of Dev Experience), Petra Grutzik (PM of Gemini Pre-training), Mehdi Ghissassi (ex-Head of Product), Karl Tuyls (ex-Research Director), Datadog: Olivier Pomel (Cofounder & CEO), Anthropic: Durk Kingma (ex-Cofounder of OpenAI), Zach Gleicher (Research PM), HuggingFace: Thomas Wolf (Cofounder & CSO), Mistral: Guillaume Lample (Cofounder & CSO), Sophia Yang (ex-Head of Dev Experience), Sentry: David Cramer (Cofounder & CPO), ARC Prize: Francois Chollet (Cofounder)
Built for production AI
Real-time protection, analytics, and optimization
for teams shipping AI at scale.



Want to deploy AI agents safely?
Control AI behavior in real time with White Circle

