Story perspectives

OpenAI Launches New Models for Enhanced Online Safety

10/30/2025

50 6

Summary

Story summary

OpenAI released the gpt-oss-safeguard-120b and gpt-oss-safeguard-20b models for online safety classification, with open weights.
They let developers view training parameters while maintaining security and help detect fake reviews and harmful content.
The models were developed with ROOST and outperform traditional classifiers on multi-policy benchmarks, though they may produce inaccurate reasoning.
The models are available under Apache 2.0 on Hugging Face and a December 8 hackathon is planned in San Francisco.

1 of 1

Story summary

OpenAI released the gpt-oss-safeguard-120b and gpt-oss-safeguard-20b models for online safety classification, with open weights.
They let developers view training parameters while maintaining security and help detect fake reviews and harmful content.
The models were developed with ROOST and outperform traditional classifiers on multi-policy benchmarks, though they may produce inaccurate reasoning.
The models are available under Apache 2.0 on Hugging Face and a December 8 hackathon is planned in San Francisco.