OpenAI open-sources teen safety policies built for gpt-oss-safeguard

OpenAI has released a set of prompt-based teen safety policies meant to help developers build age-appropriate protections into their products. The policies are designed to work directly with gpt-oss-safeguard, OpenAI's open-weight safety model, turning high-level safety requirements into classifiers that can run in real systems.

The first release covers six risk areas: graphic violent content, graphic sexual content, harmful body ideals and behaviors, dangerous activities and challenges, romantic or violent roleplay, and age-restricted goods and services. They can be used for real-time content filtering as well as offline analysis of user-generated content.

The motivation is that classifiers are only as good as their definitions. OpenAI notes that one of the biggest practical challenges for developers is writing policies that accurately capture teen-specific risks and can be applied consistently in real systems. Even experienced teams often struggle to translate high-level safety goals into operational rules, which leads to gaps in protection, inconsistent enforcement or overly broad filtering.

Shipped as prompts, the policies are meant to drop into existing workflows and to be adapted and iterated on. OpenAI says they were developed with outside organizations: Common Sense Media and everyone.ai helped shape the scope of content covered, strengthen the structure of the prompts and refine edge cases considered during evaluation.

In an accompanying comment, Common Sense Media's Robbie Torney said one of the biggest gaps in AI safety for teens has been the lack of clear operational policies developers can build from, adding that many teams start from scratch. He described the open-source policies as setting a meaningful safety floor across the ecosystem. Mathilde Cerioli, chief scientist at everyone.ai, said content policies are an important first step and open the door to work on how model behavior shapes youth-relevant risks over time.

OpenAI is explicit about the limits. The policies are a starting point, not a comprehensive definition or a guarantee of teen safety, and each application has its own risks, audiences and context. Developers are told to adapt and extend them and to pair them with product design decisions, user controls, teen-friendly transparency, monitoring systems and layered defense in depth.

The distribution channel matters too. The policies are released open source through the ROOST Model Community into the RMC GitHub repository, where developers are invited to contribute feedback, add new teen safety policies, translate them and extend them to further risk areas. gpt-oss-safeguard itself is downloadable from Hugging Face.

The release fits into a broader OpenAI push on youth protection: an updated Model Spec with Under-18 principles, product safeguards such as parental controls and age prediction, and the Teen Safety Blueprint calling for industry-wide standards. The policy pack is the open-ecosystem arm of that work.

What to watch is how the ecosystem responds: whether smaller developers actually adopt the pack, whether community contributions add risk categories beyond the initial six, and whether "policy as prompt" becomes the default way to deploy open-weight models safely.