OpenAI announced safety measures for Sora, its text-to-video generation model, in a blog post dated 23 March 2026 [source]. The company detailed content moderation systems, usage policies, and technical safeguards designed to prevent misuse of the video synthesis tool.

According to the post, Sora incorporates C2PA provenance metadata in generated videos, allowing viewers to verify AI-generated content through compatible tools. OpenAI stated that the model refuses prompts requesting depictions of public figures, copyrighted characters, and certain categories of harmful content including child safety material and graphic violence.

The safety framework includes a classifier system that reviews both text prompts and generated video frames. OpenAI reported that approximately 10 percent of generation attempts are blocked by these filters. The company also described a feedback mechanism allowing users to report outputs that bypass safety systems.

OpenAI disclosed that Sora access remains limited to approved creators and red team researchers. The company did not specify a timeline for broader public release, stating only that expanded access would follow "additional safety testing and refinement."

The announcement follows earlier provider disclosures about generative model safety challenges. OpenAI noted that Sora's training data includes "publicly available and licensed content," but did not detail specific licensing arrangements or data sources.

The post acknowledged ongoing challenges in preventing adversarial prompt engineering, stating that "no safety system is perfect" and that the company expects users to discover edge cases that circumvent current protections. OpenAI committed to iterative updates based on observed misuse patterns.