Idea
An open-weight constitutional classifier that filters unsafe prompts and outputs across languages with fewer false alarms and missed harms.
Research Paper
Core Innovation
This paper introduces HaloGuard 1.0, a constitutional-classifier model that leverages a detailed natural-language constitution to generate exhaustive paired counterfactual training data across 46 languages. It achieves superior multilingual prompt-safety performance at a fraction of the size of existing models by balancing false positives and false negatives and treating language as a surface form rather than an adversarial factor.
Why It Matters
AI systems handle more languages and more sensitive content every quarter, and existing safety guards miss harms or throw too many false positives. Built on fine-tuned Qwen3.5 checkpoints (0.8B and 4B) trained as generative classifiers, at under a billion parameters this one outperforms guard models 30 times its size. Nobody pays for a classifier alone. Lakera, HiddenLayer, and Robust Intelligence built businesses on the audit logs, compliance packaging, and red-teaming wrapped around a detector like this.
Market Size (TAM)
$2–10B TAM for AI content safety and moderation; $500M–$1B SAM from AI platforms, enterprises, and regulators. Driven by increasing AI adoption and regulatory pressure for safe AI outputs.
Potential Customers & Pain Points
- AI platform providers – Need reliable multilingual content safety
- Enterprises deploying chatbots – Require low false positives to maintain user trust
- Regulators and compliance teams – Demand transparent and auditable safety models
- AI developers – Seek efficient models for prompt safety without large compute costs.
Business Model
Open-weight model release with potential revenue from enterprise support, customization services, and integration partnerships.
Competitive Landscape
- OpenAI Moderation API
- Google Perspective API
- Anthropic's Constitutional AI
- Cohere Safety Models
Implementation Challenges
- Integration complexity with diverse AI systems
- Continuous adaptation to evolving adversarial attacks
- Benchmark mislabeling affecting evaluation accuracy
Validation Strategy
- Deploy in real-world AI platforms for multilingual content moderation
- Conduct adversarial red-teaming to identify and fix vulnerabilities
- Benchmark against leading commercial and open-source safety models
- Gather user feedback on false positive/negative rates in production
Research Paper Overview
HaloGuard 1.0: An Open Weights Constitutional Classifier for Multilingual AI Safety
Summary
HaloGuard 1.0 is an open-weights constitutional classifier that delivers state-of-the-art multilingual prompt-safety detection with significantly smaller model sizes. It uses a natural-language constitution of policies to generate balanced synthetic data across 46 languages, achieving high precision and low false-positive and false-negative rates on multiple benchmarks. The model continuously improves through adversarial red-teaming and is released openly for broad adoption.