Public Access to Open AI's Sora Video Generator just Leaked...

OpenAI’s Sora video generator appears to have leaked into public access through a Hugging Face space, sparking both excitement over new AI video demos and backlash over how the access was obtained. The leaked material is framed as “Sora generations” with visible OpenAI branding and a watermark, and early examples—shared via Twitter—range from walking-through-a-portal scenes and smooth city footage to more character-driven clips like a cat chasing a mouse. The access is described as free and tied to an open letter accusing OpenAI of “artwashing,” where artists were used as unpaid testers and then asked to share outputs for marketing value.

The open letter, addressed to “corporate AI overlords,” claims artists were granted early access under promises of being “early testers, red teamers and creative partners,” but were instead pulled into a PR pipeline. It argues that artists are effectively “unpaid R&D,” with compensation limited to the possibility of a small selection of films being screened after a competition—while OpenAI gains substantial marketing and validation from the resulting content. The letter also alleges that outputs require OpenAI approval before sharing, and it criticizes the imbalance between the scale of unpaid contributions and the value of OpenAI as a 150 billion company.

Alongside the ethics dispute, the leaked access is portrayed as limited and technical: the generations are said to come from a “Sora turbo” variant capped at 10 seconds per generation at 30 FPS, with output up to 1080p. The repeated emphasis on “turbo” suggests a larger, more capable Sora model remains undisclosed. The transcript also notes that the Hugging Face space became overwhelmed and that OpenAI may have pressured Hugging Face to remove or disable the generation function, even if the space link still exists.

Quality assessments in the transcript are mixed but generally favorable for motion consistency. Clips are praised for stable camera movement, coherent physics (shirts, tails, fur, reflections), and fewer “mushing” errors in body parts—especially in animal and walking scenes. Still, some examples are described as weaker, including glitches in spatial understanding (a robot-like figure with duplicated head/arm issues), blurry or inconsistent eating scenes (spaghetti/burger attempts), and occasional oddities like a cat’s foot/tail behavior that later corrects.

The fallout extends beyond the leak itself. The transcript claims OpenAI’s Discord reportedly banned Sora as a topic, and it frames the situation as a reputational hit for OpenAI—especially given earlier delays and the company’s shift from nonprofit roots toward a more closed, for-profit posture. While the transcript condemns the leak as potentially illegal and risky for those involved, it also argues that the incident reveals how much remains hidden behind closed doors—leaving open the question of what the full, larger Sora model can do and whether it will be released publicly on a timeline that matches expectations.