Skip to content

Latest commit

 

History

History
75 lines (66 loc) · 6.7 KB

attack-surface.md

File metadata and controls

75 lines (66 loc) · 6.7 KB

Attack Surface Analysis for commaai/openpilot

Attack Surface: CAN Bus Injection (via openpilot)

  • Description: Unauthorized modification of messages on the vehicle's CAN bus through the compromised openpilot device.
  • openpilot Contribution: openpilot's required CAN bus access for core functionality creates the direct pathway. It acts as a potentially compromised bridge.
  • Example: An attacker, having gained RCE on the openpilot device, sends CAN messages to disable the brakes.
  • Impact: Loss of vehicle control, unintended actions (acceleration, braking, steering), disabling safety systems, vehicle damage, injury/death.
  • Risk Severity: Critical
  • Mitigation Strategies:
    • Developers:
      • Secure Boot: Prevent unauthorized firmware execution.
      • Strict CAN Message Filtering: Implement rigorous input validation and filtering (both inbound and outbound) on the CAN bus. Whitelist allowed messages; blacklist/reject all others.
      • CAN Bus Intrusion Detection (IDS): Monitor for anomalous CAN traffic patterns.
      • Hardware Security Module (HSM): Protect cryptographic keys and enable secure CAN message signing/verification.
      • Code Hardening (CAN Interface): Focus secure coding practices specifically on the CAN communication code.
    • Users:
      • Physical Device Security: Prevent unauthorized physical access to the openpilot hardware.

Attack Surface: Remote Code Execution (RCE) on openpilot Device

  • Description: An attacker gains the ability to execute arbitrary code on the openpilot device itself.
  • openpilot Contribution: openpilot's software complexity, reliance on external libraries (C++, Python, OS components), and network connectivity (Wi-Fi, cloud) create the vulnerability surface.
  • Example: Exploiting a buffer overflow in openpilot's image processing code via a crafted image sent over Wi-Fi.
  • Impact: Complete device compromise, leading to CAN bus injection (see above), data exfiltration, and potential lateral movement within the vehicle's network.
  • Risk Severity: Critical
  • Mitigation Strategies:
    • Developers:
      • Secure Coding Practices (OWASP): Adhere rigorously to secure coding guidelines.
      • Memory Safety: Prioritize memory-safe languages (e.g., Rust) or robust memory management in C/C++.
      • Regular Vulnerability Scanning: Employ static and dynamic analysis tools.
      • Dependency Management (SBOM): Track, vet, and update all dependencies.
      • Sandboxing: Isolate openpilot components (sensor processing, control, networking) to contain breaches.
      • Secure Update Mechanism: Implement robust, cryptographically verified updates.
      • Principle of Least Privilege: Minimize privileges for openpilot processes.
    • Users:
      • Prompt Software Updates: Install updates immediately when available.
      • Strong Wi-Fi Security: Use a strong, unique password for the connected Wi-Fi network.

Attack Surface: Sensor Spoofing Leading to Incorrect Model Decisions (openpilot-Specific)

  • Description: Manipulating sensor data (camera, radar) to cause openpilot's specific machine learning models and control algorithms to make incorrect driving decisions.
  • openpilot Contribution: This is not general sensor spoofing; it's about exploiting the unique way openpilot processes and reacts to sensor data. The vulnerability lies in openpilot's specific model architecture, training data, and sensor fusion logic.
  • Example: Projecting a fake lane marking that specifically triggers a misclassification in openpilot's lane detection model, causing a lane departure.
  • Impact: Incorrect lane keeping, false obstacle detection/avoidance, incorrect speed control, leading to potential accidents.
  • Risk Severity: High
  • Mitigation Strategies:
    • Developers:
      • Sensor Fusion with Redundancy: Cross-validate data from multiple, independent sensors.
      • Adversarial Training: Train models with adversarial examples to improve robustness.
      • Anomaly Detection (Sensor Data): Detect physically impossible or highly improbable sensor readings.
      • Contextual Awareness: Integrate map data, time of day, and other contextual information to validate sensor input.
      • Out-of-Distribution (OOD) Detection: Identify when sensor inputs deviate significantly from training data.
    • Users:
      • Maintain Sensor Cleanliness: Ensure sensors are unobstructed.
      • Situational Awareness: Remain attentive and ready to take over control.

Attack Surface: Model Poisoning / Data Poisoning (Targeting openpilot's Models)

  • Description: Tampering with the training data used to create openpilot's machine learning models, causing them to misbehave in specific, attacker-controlled ways. This targets the specific models used by openpilot.
  • openpilot Contribution: The vulnerability exists because openpilot relies on machine learning and updates to those models. The attack targets the integrity of openpilot's training pipeline.
  • Example: Subtly modifying a dataset of driving images to cause openpilot's lane-keeping model to consistently drift left under specific, rare conditions.
  • Impact: Subtle but potentially dangerous misbehavior of openpilot, difficult to detect until triggered, leading to accidents in specific scenarios.
  • Risk Severity: High
  • Mitigation Strategies:
    • Developers:
      • Data Provenance and Integrity: Securely track and verify the origin and integrity of all training data.
      • Data Sanitization: Thoroughly inspect and clean training data.
      • Robust Training Techniques: Employ methods less susceptible to data poisoning (adversarial training, differential privacy).
      • Model Monitoring: Continuously monitor deployed model performance for anomalies.
      • Red Team Exercises (Data Poisoning): Simulate poisoning attacks to test defenses.
      • Secure Model Storage: Protect trained models from unauthorized access.
    • Users: (Limited direct mitigation; relies on developer actions)
      • Install Only Official Updates: Ensure updates come directly from comma.ai and are verified.