Ali Eslami

AI-Controlled Systems Vulnerable to Stealthy Gain Manipulation Without Triggering Safety Checks

Autonomous AI agents integrated into cyber-physical systems create a novel attack surface through parameter update pathways absent in traditional control architectures. Attackers can manipulate feedback gain matrices to destabilize closed-loop dynamics while evading residual-based security monitors, with a single gain update capable of altering eigenvalue placement across the entire system. Critically, a gain replacement can preserve overall stability while producing dangerous transient amplification that exceeds safe operating limits—meaning stability verification alone is insufficient for security assurance.

Stealthy Adversarial Attacks on CAV Platoons Can Hijack Trajectories While Evading Anomaly Detection

This paper reframes vehicle platoon security from a defender-centric to an attacker-centric perspective, formally characterizing how adversaries can covertly steer connected autonomous vehicle (CAV) platoons toward attacker-chosen trajectories without triggering anomaly detection. The authors derive feasibility conditions for such stealthy attacks and show their dependence on communication topology and control protocol choices. By quantifying the resources an attacker needs, the work exposes structural vulnerabilities in current platoon architectures and provides a foundation for designing more resilient countermeasures.

Security Risks of Agentic Vehicles: A Systematic Analysis of Cognitive and Cross-Layer Threats

A Control-Theoretic Foundation for Agentic Systems

Covert Attacks Pose Greatest Threat to Vehicle Lateral Control — Sensor Selection Is a Key Defense Lever

This paper applies a system-theoretic framework to analyze three classes of stealthy cyber-attacks — replay, zero dynamics, and covert — against vehicle lateral dynamics modeled via the bicycle model. The central finding is that attack feasibility and impact are strongly coupled to sensor/output configuration: output selection can eliminate unstable zero dynamics, neutralizing zero dynamics attacks entirely. Covert attacks, which require coordinated actuator and sensor manipulation plus deep system knowledge, are the most dangerous class, enabling sustained, undetected lateral state deviation. Actuator and tire saturation introduce attack-class-dependent effects on both stealth and effectiveness, validated through CarSim-Simulink co-simulation.

Multimodal Real-Time AI Co-Clinician Approaches PCP Performance in Telemedicine Simulations

Researchers from Google DeepMind and collaborators introduce "AI co-clinician," a dual-agent conversational AI system built on Gemini that processes continuous audio-visual streams from live patient encounters to support real-time clinical decision-making. Evaluated against primary care physicians (PCPs), GPT-Realtime, and a baseline agent across 120 simulated telemedicine encounters, the system approached PCP-level performance on management plans and differential diagnosis while significantly outperforming GPT-Realtime on all general criteria. Key gaps persist in physical examination reasoning and disease-specific assessment, reinforcing the authors' conclusion that high-stakes diagnostic AI is best deployed in a triadic, human-collaborative model rather than autonomously. The study also demonstrates that text-only AI evaluation frameworks fail to capture the full complexity of real medical consultation.