Comma AI — agentic threat model

Uses end-to-end deep learning models for vision and vehicle path planning. Primary threats include adversarial physical attacks (e.g., adversarial stickers on road signs or lane markings) that can trick the model into dangerous driving maneuvers, and model poisoning via corrupted driving logs.

Relies heavily on crowdsourced real-world driving data for training. Threats include data poisoning of the training pipeline by submitting malicious driving logs, and privacy/exfiltration risks associated with uploading high-resolution video and GPS logs from user vehicles.

The openpilot software orchestrates inputs from cameras/sensors and translates them into CAN bus control messages. Vulnerabilities in this orchestration layer (written in C++/Python) could allow memory corruption or logic bypasses, leading to unauthorized tool execution (unintended steering or braking commands).

Deployed on custom hardware (comma 3X) running a custom operating system. Threats include local privilege escalation to gain root access, insecure over-the-air (OTA) update mechanisms, and physical port exploitation allowing attackers to flash malicious firmware directly to the device.

Features an active driver monitoring system (DMS) to ensure human attention. Threats include evaluation gaming (spoofing the driver-facing camera to bypass attention checks) and insufficient logging/fail-safes when the system encounters out-of-distribution road scenarios.

As an aftermarket, open-source ADAS, it operates in a complex regulatory landscape. Key threats include lack of formal automotive safety certifications (like ISO 26262) and potential compliance gaps regarding regional autonomous driving regulations and liability frameworks.

Not certain from the listing — Comma AI does not explicitly detail multi-agent coordination or a decentralized agent marketplace. However, risks in this layer would involve fleet-wide cascading failures if a corrupted model update is pushed to all connected vehicles simultaneously.