AI Security Engineer, Simulation & Adversarial Validation - Physical Team

Join Lenovo’s Physical AI Security team to develop security technologies that protect how robots and autonomous systems perceive, decide, and act. You will investigate how attackers can manipulate AI-driven behavior and build runtime defenses that detect compromised inputs, constrain unauthorized actions, and support controlled recovery. 

This role combines security research with robotics, controls, multimodal AI, predictive modeling, and reinforcement learning. 

Candidate must be willing to work onsite three days a week in our Tel-Aviv, Israel office. 

Job Responsibilities: 

• Investigate adversarial manipulation of VLA/VLM models, learned policies, and world models, including malicious instructions, manipulated observations, forged robot state, and compromised decision context. 

• Analyze attack surfaces and trust boundaries across perception, planning, control, and actuation. Reproduce attacks in simulation and controlled robot experiments to evaluate their impact. 

• Build learned dynamics and world models that evaluate proposed actions, anticipate harmful physical outcomes, and estimate whether intervention or recovery remains feasible. 

• Combine analytical dynamics, neural residual models, and state estimation to identify physically inconsistent observations, unexpected motion, and deviations from authorized behavior. 

• Implement deterministic, real-time C++ and ROS 2 controls that validate AI-generated commands, enforce action permissions and physical limits, and block or constrain commands that violate policy. 

• Apply model-based, offline, and constrained reinforcement learning to action evaluation, policy shielding, and recovery under manipulated observations or degraded operating conditions. 

• Use uncertainty calibration, distribution-shift detection, and adversarial testing to characterize detection coverage, false positives, bypass conditions, and intervention latency. 

• Collaborate with university laboratories and graduate researchers, supervise research projects, and contribute to publications in robotics, AI security, and Physical AI cybersecurity. 

Minimum Requirements: 

• Advanced C++, Python, Linux, and ROS/ROS 2, with hands-on experience integrating AI and control software with physical robots. 

• Experience in cybersecurity, adversarial machine learning, or security research involving autonomous, embedded, or cyber-physical systems. 

• Ability to translate threat models into reproducible attacks, measurable security requirements, and practical defenses. 

• Strong PyTorch experience, including training and evaluating multimodal, transformer, diffusion-policy, or sequence models, with hands-on TensorRT/CUDA experience for inference optimization and deployment. 

• Strong knowledge of robot dynamics, feedback control, MPC, system identification, and state estimation, with an understanding of how corrupted inputs and timing disruptions affect control behavior. 

• Experience with VLA/VLM action representations, learned policies, world models, or imitation learning. 

• Experience with model-based, offline, or constrained reinforcement learning, policy shielding, or recovery estimation. 

• Strong understanding of uncertainty calibration, OOD detection, and model evaluation, including the limitations of using these methods to identify adversarial behavior. 

Preferred Requirements: 

• M.Sc. or Ph.D. in a relevant field, with research experience in robotics, AI, controls, or cybersecurity, is strongly preferred.