OpenAI Releases GPT-OSS-SafeGuard Models for Customizable AI Safety

OpenAI has released a new safety-focused family of open-source models, GPT-OSS-SafeGuard-20B and GPT-OSS-SafeGuard-120B, designed for security classification tasks and enabling developers to define their own safety boundaries. Building on earlier open-source releases, GPT-OSS-20B and GPT-OSS-120B, the SafeGuard models interpret developer-defined safety policies during inference rather than relying on static pre-trained constraints.

They can perform safety classification on user messages, chat segments, or entire conversation histories, depending on the application’s requirements, with policies applied at inference time so strategies can be refined over time. Each model accepts two inputs—a policy and the content to be analysed—and outputs both the classification result and a reasoning trace explaining the decision.

According to OpenAI, these models are particularly suited when risks are evolving, the domain is too sensitive for small classifiers, or there is insufficient labeled data, though they are computationally heavier and slower than traditional classifiers. The SafeGuard models are released under the Apache 2.0 licence and can be downloaded from OpenAI’s official repository on HuggingFace.