Seedance 2.0 AI Video — agentic threat model

Uses proprietary multimodal foundation models for text-to-video and audio-to-video generation. Primary threats include adversarial prompt injection to bypass safety filters (generating NSFW or non-consensual deepfakes) and model extraction/stealing of their proprietary video generation weights.

Not certain from the listing — No details are provided regarding data storage, vector databases, or training pipelines. However, the system processes sensitive user-uploaded reference images, videos, and audio, which are vulnerable to data exfiltration, unauthorized access, or potential poisoning if user uploads are used for continuous model fine-tuning.

Not certain from the listing — The orchestration framework for multi-shot storytelling and @-reference multimodal control is not specified. Vulnerabilities could include insecure state handling across multi-shot generation steps or prompt injection manipulating the generation pipeline's logical flow.

Not certain from the listing — Hosted as an online closed-source SaaS. The heavy GPU compute required for video rendering makes it a prime target for denial-of-service (DoS) attacks, resource exhaustion, or unauthorized GPU mining if infrastructure sandboxing is weak.

Not certain from the listing — No observability, logging, or automated guardrails are detailed. A lack of robust input/output moderation filters represents a significant gap, allowing users to generate harmful synthetic media or bypass copyright restrictions.

Not certain from the listing — No compliance certifications (such as SOC2 or GDPR) are mentioned. The capability to perform lifelike lip-syncing of 'real human video' in 8+ languages introduces severe compliance risks under deepfake and biometric regulations (e.g., EU AI Act requirements for synthetic media labeling).

Not certain from the listing — The agent operates as a standalone horizontal tool without an active multi-agent ecosystem or marketplace integration described in the directory listing.