Pony.ai — agentic threat model

Not certain from the listing — The specific deep learning or foundation models powering the perception and prediction pipelines are proprietary. Threats include adversarial physical attacks (e.g., modified stop signs) and model evasion techniques that could blind the vehicle's perception.

Not certain from the listing — The data operations layer likely processes massive streams of sensor, LiDAR, and camera data for training and HD mapping. Threats include training data poisoning and simulation data manipulation, which could introduce latent driving vulnerabilities.

The 'Virtual Driver' acts as the core agent framework, orchestrating localization, route planning, and real-time control. Threats include planning logic bypasses or sensor-fusion tool misuse, leading to erratic physical vehicle behavior.

Deployment occurs on safety-critical, edge-compute hardware within physical vehicles. Threats include remote code execution via telematics, OTA update compromise, and physical access exploits targeting the vehicle's internal CAN bus.

With millions of kilometers driven, real-time observability and safety guardrails are critical. Threats include sensor drift, blind spots in anomaly detection, and the failure of safety-critical fallback systems to disengage the virtual driver during a cyber-physical attack.

Not certain from the listing — While automotive safety standards (like ISO 26262 or ISO/SAE 21434) are standard in the industry, the listing does not explicitly detail compliance frameworks, identity management, or access control policies for vehicle fleet operations.

The agent ecosystem involves fleet management, robotaxi/robotruck dispatch networks, and potential V2X (Vehicle-to-Everything) communications. Threats include cascading failures from compromised dispatch servers or rogue vehicle agents broadcasting false traffic data.