Vidan.ai — agentic threat model

Not certain from the listing — likely utilizes specialized computer vision models (such as CNNs or ViTs) for object, facial, and hazard detection. Primary threats include physical adversarial attacks (e.g., adversarial patches that blind the model to weapons or intruders) and model evasion.

Not certain from the listing — processes real-time video streams and potentially stores reference data like facial templates or license plate watchlists. Threats include unauthorized access to sensitive video feeds, data exfiltration of private surveillance data, and poisoning of reference databases.

Not certain from the listing — orchestration appears to be a pipeline connecting video ingestion to alerting and reporting mechanisms. Threats include manipulation of alert thresholds or suppression of critical security events.

Not certain from the listing — likely deployed on edge devices near cameras or in a centralized cloud. Threats include compromise of edge hardware, interception of unencrypted RTSP video streams, and unauthorized access to local network segments.

Not certain from the listing — requires continuous monitoring to detect camera tampering, occlusion, or model drift due to environmental changes (e.g., lighting). Threats include silent failures where the system stops detecting hazards without alerting administrators.

Not certain from the listing — facial recognition and license plate tracking trigger strict regulatory requirements (e.g., GDPR, EU AI Act high-risk categorization). Threats include non-compliance, lack of user consent mechanisms, and insufficient audit logging of who accessed surveillance data.

Not certain from the listing — no multi-agent or marketplace ecosystem is described. Threats are limited to downstream integrations with third-party physical security systems or notification APIs failing to deliver critical alerts.