Underage sexual content, self-harm information targeted by new OpenAI open source directives

OpenAI has announced new open source security guidelines for developers, which aim to introduce a number of new security policies.

The information-based safety pack includes model guidance on common youth risks, recommendations for developmental content, and age-appropriate guidelines on topics such as self-harm, sexual content and romantic role-playing, dangerous trends or viral challenges, and harmful physical norms.

OpenAI said it’s a more robust alternative to the high-level guidelines previously provided, formatted as pluggable instructions for AI applications.

OpenAI added new under-18 goals to its Model Spec in December. A few months ago, the company released gpt-oss-safeguard, an open-source logic model designed to help developers implement security scenarios and distinguish between safe and unsafe content. Unlike traditional security classification processes, gpt-oss-safeguard can be fed into the platform’s security policies directly, and includes the intent of the policy as it classifies the appropriate output.

But “even experienced teams often struggle to translate high-level security principles into concrete, actionable rules, especially since they require both subject matter expertise and deep AI knowledge,” OpenAI said in its latest press release. “This can lead to gaps in protection, inconsistent enforcement, or overly broad filtering. Clear, well-planned policies are a critical foundation for effective security programs.”

The additional developer pack was designed in collaboration with the non-profit Common Sense Media and everyone.ai.

Experts have warned parents about chatbot overexposure of vulnerable teenagers and even young children, as AI companies try to get a handle on the effects of their models on users’ mental health. Last year, OpenAI was sued by the parents of toddler Adam Raine in the industry’s first wrongful death lawsuit, with the Raine family claiming a combination of ChatGPT sycophancy and lax security policies caused their son’s suicide. The company has denied allegations of wrongdoing and in response has strengthened its mental health and youth safety features, including age verification. However, third-party developers who license OpenAI models have struggled to maintain the same level of safety measures, including for children’s AI-powered toys.

The lawsuit against OpenAI followed several lawsuits against the controversial platform Character.AI and set the stage for the latest wrongful death lawsuit filed by OpenAI rival Google and its Gemini AI backer.

Companies across industry, technology and social media are facing legal challenges regarding the long-term impact of their products on users. Last month, Instagram CEO Adam Mosseri and Meta chief Mark Zuckerberg testified in front of a jury in a watershed trial that puts the social networks on trial for their allegedly addictive design principles. A decision has not been reached.

OpenAI said its new security package is not a “comprehensive or final definition or guarantee of security for young people.” Robbie Torney, head of AI and digital analytics for Common Sense Media, said the new policies could create a “profitable security layer in the ecosystem,” filling the AI ​​security gap that has been exacerbated by a lack of developer policies.

Developers can download the OpenAI security model from Hugging Face and access its new information pack on GitHub.

Disclosure: Ziff Davis, Mashable’s parent company, in April filed a lawsuit against OpenAI, alleging that it infringes Ziff Davis’ copyrights in training and using its AI programs.