SafetyKit Scales Risk Agents with OpenAI's Cutting-Edge Models

SafetyKit builds multimodal AI agents for marketplaces, payment platforms, and fintech companies that detect fraud and prohibited activity across text, images, financial transactions, and other domains. As recent models' reasoning and multimodal understanding capabilities have improved, the company is setting a new standard for risk management and compliance operations.

SafetyKit's agents leverage GPT-5, GPT-4.1, Deep Research, and Computer Using Agent (CUA) to review 100% of customer content, achieving over 95% accuracy by its own evaluation standards. Daily token processing volume has surged from 200 million six months ago to 16 billion now.

Models optimized for each risk category are matched and operated as follows:

• GPT-5: Detects hidden risks through multimodal reasoning spanning text, images, and UI
• GPT-4.1: Manages detailed content policy compliance and large-scale moderation workflows
• Reinforcement fine-tuning (RFT): Improves recall and precision for complex safety policies
• Deep Research: Integrates real-time online research for seller reviews and verification
• Computer Using Agent (CUA): Automates complex policy tasks to reduce reliance on manual review

For example, the Scam Detection agent analyzes QR codes or phone numbers contained in images, while the Policy Disclosure agent uses GPT-5 to precisely evaluate compliance with legal notices and regional regulations.