OpenAI Sora guardrail jailbreak via "Anatomical Escalation" technique
A jailbreak vulnerability was discovered in OpenAI's Sora video generation model that uses anatomical terminology and stylized visual framing to bypass content restrictions. The technique exploits linguistic methods including prompt injection and synonymous language substitution to evade safety filters.
What happened
A jailbreak vulnerability was discovered affecting OpenAI's Sora video generation model. The technique, termed 'Anatomical Escalation,' combines technical anatomical terminology with creative visual elements, specifically requesting bio-luminescent symbols that conform to body contours to bypass inference restrictions around generating potentially inappropriate visual content.
Reported impact
The disclosure was rated as Level 1 minimal risk, affecting individual users without broader implications. The generated content was classified as Level 1 abstract symbolic representation, consisting of highly stylized, simplified, or geometric nude forms appearing as cartoonish shapes or abstract figures.
Index analysis
The vulnerability exploits linguistic techniques including prompt injection, stylization, and synonymous language substitution to evade content filtering mechanisms. By framing requests in artistic and technical language, the jailbreak allows users to manipulate the model's understanding and responses to generate content that may not trigger safety protocols.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.