LLM-Enabled Robot Threat Modeling Study
A new study from arXiv (2604.27267) models cybersecurity threats in LLM-enabled autonomous robotic systems using an edge-cloud architecture. The research applies STRIDE-per-interaction analysis across six boundary-crossing points, categorizing threats into Conventional Cyber, Adversarial, and Conversational types. It finds these categories converge at boundary crossings and traces three cross-boundary attack chains from external inputs to physical actuation. The work addresses a gap in prior studies that treated robotic cybersecurity, adversarial perception, and LLM safety separately.
Key facts
- arXiv paper 2604.27267
- Models LLM-enabled autonomous robot in edge-cloud architecture
- Uses hierarchical Data Flow Diagram
- Applies STRIDE-per-interaction analysis
- Six boundary-crossing interaction points
- Three-category taxonomy: Conventional Cyber, Adversarial, Conversational
- Traces three cross-boundary attack chains
- Addresses gap in unified threat modeling
Entities
Institutions
- arXiv