The identified vulnerability in OpenAI’s Sora 2 demonstrates how audio can be used effectively to unearth hidden system prompts, revealing serious security implications for multimodal AI systems. Researchers from Mindgard discovered that by transcribing audio outputs, they could compile a nearly complete version of the model’s internal rules and boundaries, surpassing previous visual methods that failed to accurately convey information.
The implications of this vulnerability go beyond mere academic interest; they signify potential risks whereby leaked prompts may enable users to craft malicious requests to circumvent safety measures. Mindgard urges a reconsideration of how these internal instructions are treated, suggesting they be protected as configuration secrets akin to firewall rules. This vulnerability signals a pressing need for ongoing vigilance in the field of AI, especially as the landscape evolves with increasing deepfake and disinformation threats, necessitating stronger protective measures as multimodal systems become more commonplace.
👉 Pročitaj original: Cyber Security News
