Source-linked AI summary
On the security and privacy of LLMs in Mobility
Mauro Conti, Lorenzo Perinello, Umberto Salviati
TL;DR
Mobility LLM research lacks systematic security-focused assessment, despite security and privacy concerns in intelligent transport systems. This paper surveys the literature and evaluates it against AI Act-derived criteria; it finds GPT and Llama dominate while security measures remain uncommon.
Problem
Method
The paper surveys LLM mobility research and assesses security, privacy, and reliability, translating EU AI Act provisions into 9 technical evaluation criteria.
Source: §1 · p. 2Results
GPT and Llama account for more than half of surveyed papers, while only five discuss security and two incorporate security mechanisms.
Source: §5 · p. 8Takeaways & Limitations
The authors call for compliance frameworks and cybersecurity mechanisms embedded in Intelligent Transportation Systems across the AI lifecycle.
Source: §5 · p. 8The assessment covers algorithmic architectures rather than commercial systems and excludes operational and administrative deployment duties.
Source: §4.3.1 · p. 7
Abstract
from arXiv · showhide
The mobility sector is undergoing a paradigm shift driven by advances in Generative Artificial Intelligence. With a global market valued at approximately 2.9 trillion dollars annually, considering only cars, the integration of these technologies has the potential to impact more than 1.5 billion vehicles worldwide. As Large Language Models (LLMs) are increasingly adopted in mobility, concerns about cybersecurity, privacy, and reliability emerge. Accordingly, this paper surveys current applications and assesses these challenges. Since the European AI Act classifies transportation AI as high risk, we derive nine technical classes from its requirements to assess current research and future deployments. Our findings show that research mainly studies GPT and Llama models (over 50\% of reviewed works) and traffic applications while largely neglecting security, privacy, and reliability. This gap extends to AI Act compliance: among 35 reviewed works, only one includes a partial vulnerability assessment and one a partial risk management system. We identify a clear gap between strong optimization performance and regulatory adherence, suggesting compliance is limited less by technology than by a focus on static performance over lifecycle safety, and underscoring an urgent need for security-by-design in safety-critical intelligent transportation systems.
1 Introduction
Mobility systems increasingly integrate AI and LLMs, while intelligent transport systems face security, privacy, and reliability challenges. The paper surveys these concerns and assesses mobility research against technical criteria derived from the European AI Act.
1 Introduction
- Intelligent Transport Systems integrate information and communication technologies into transport, but also face cyber threats, data protection issues, and reliability risks.The paper describes security as fundamental to the design, deployment, and management of modern mobility systems.Source: §1 · p. 1
- Generative AI adoption in transport has focused on performance and operational capabilities, while the authors identify security-by-design as systematically overlooked.The authors state that rapid deployment of LLMs and autonomous agents creates a threat surface exploitable by malicious actors.Source: §1 · p. 1
- The paper addresses the lack of a systematic security-focused study of LLMs in mobility by mapping deployment gaps, threats, and AI Act requirements.It also investigates security, privacy, and reliability across the surveyed works.Source: §1 · p. 2
- The paper translates EU AI Act provisions into 9 technical evaluation criteria for AI transportation systems.The criteria are used to assess the works investigated in the survey.Source: §1 · p. 2
2 Related Work
Prior surveys examine LLM applications across transportation, but the cited literature does not systematically address the security challenges associated with adopting LLMs in mobility.
2 Related Work
- LLM security research identifies jailbreaking and prompt injection as threats; indirect prompt injection can arrive through untrusted documents or web pages [70] [16].Jailbreaking bypasses alignment guardrails, while prompt injection uses adversarial text to override system constraints.Source: §2 · p. 2
- Existing transportation surveys cover applications including traffic prediction, signal optimization, autonomous driving, and vehicle and pedestrian detection [45].Other surveys examine models, methods, datasets, traffic management, predictive maintenance, vehicular communications, planning, and safety.Source: §2 · p. 2
- The cited mobility surveys do not systematically address security or related adoption challenges, leaving security largely overlooked in this literature.The paper positions its security-focused analysis as addressing this gap.Source: §2 · p. 2
3 Survey
The survey searches and classifies LLM mobility research, then assesses how reviewed works address security, privacy, reliability, and reproducibility. Its findings show concentrated model use, sparse security measures, and limited code availability.
3 Survey
- The survey examines models, tasks, security, privacy, reliability, and reproducibility in urban mobility research.Its findings are intended to describe the landscape and identify research gaps.Source: §3 · p. 2
3.1 Methodology
- The Web of Science search returned 125 publications, of which 35 remained for full-text analysis after screening.Google Trends relative search interest was used as a proxy for public attention and visibility, not actual usage.Source: §3.1 · p. 2
3.2 Key elements inspected
- The review records model functionality and task, and classifies security, privacy, and reliability as discussed, implemented, or not addressed.The functionality categories follow a taxonomy covering information processing, knowledge encoding, component generation, and decision facilitation.Source: §3.2 · p. 3
3.3 Findings
- The findings overview compares commonly adopted models and application tasks with coverage of security, privacy, reliability, and reproducibility.The section frames these dimensions as aspects considered in current research.Source: §3.3 · p. 3
3.3.1 Models.
- GPT appeared in 19 reviewed papers and Llama in 11, with GPT accounting for more than half of the reviewed papers.Figure 1 presents the distribution of model families across the reviewed papers.Source: Figure 1 · p. 3, §3.3.1 · p. 3
3.3.2 Functionality.
- Decision facilitators were the most common functionality, appearing in 25 of 35 studies (71.4%).Information Processor appeared in 18 studies (51.4%), Knowledge Encoders in 11 (31.4%), and Functionality C in one; A and D co-occurred in 10 studies.Source: §3.3.2 · p. 3
3.3.3 Security, Privacy, Reliability.
- Security measures appeared in only two studies, privacy was usually discussed only at a high level or as future work, and six studies incorporated reliability mechanisms.Only explicitly considered data privacy in dataset selection; reliability mechanisms included data-authenticity measures and human intervention [73], output validation, and output comparison and consistency checks [43].Source: §3.3.3 · p. 3
3.3.4 Code Availability.
- 13 of the 35 investigated papers made code available through a repository.The paper identifies open-source availability as important for reproducibility and security-related research.Source: §3.3.4 · p. 3
4 AI Act in Transportation Reserach
The section maps transportation AI Act obligations to technical criteria and assesses how reviewed mobility research addresses them. It finds strong functional-performance focus alongside sparse security and compliance safeguards.
4.1 Risk Classification in Transportation Systems
- Road-traffic management and transport operations are classified as high-risk because system failure could endanger public health and safety.The section connects that classification to LLMs and agentic architectures used for mobility tasks such as traffic management and vehicle dispatching.Source: §4.1 · p. 5
4.2 Compliance Framework Evaluation Criteria
- The framework translates high-level AI Act requirements into transportation-specific evaluation criteria, including adversarial testing, cybersecurity, lifecycle risk management, and post-market monitoring.Other criteria address accuracy and robustness, transparency and copyright, human oversight, data governance, and compliance documentation.Source: §4.2 · pp. 5–6
4.3.3 Empirical Evaluation.
- Accuracy & Robustness averages μ=2.30 (σ=1.21), versus Vulnerability Evaluation at μ=1.11 (σ=0.51) and Risk Management System at μ=1.09 (σ=0.37).The assessment reports pronounced divergence between functional performance and security compliance.Source: Table 2 · p. 6, §4.3.3 · p. 7
4.3 Compliance Assessment of the Reviewed Work
- The assessment evaluates 35 selected papers against nine technical criteria derived from AI Act mandates.The paper describes its dataset as 280 evaluation points, calculated as 35 papers × 8 criteria.Source: §4.3 · p. 7
4.3.1 Scope Selection.
- The scope assesses requirements within researchers’ influence and excludes operational and administrative duties assigned to commercial deployment.Included areas cover vulnerability evaluation, risk management, transparency and copyright, and accuracy and robustness.Source: §4.3.1 · p. 7
4.3.2 Evaluation Metric and Likert Scale.
- A five-point Likert scale rates each requirement from omission or acknowledgment through assessment and implementation to integrated compliance workflows.Static, one-time testing corresponds to “Assessed,” while full integration requires automated compliance workflows.Source: §4.3.2 · p. 7
5 Conclusion and Future Directions
The survey finds that mobility LLM research gives limited attention to security, privacy, and reliability, alongside weak AI Act alignment. It calls for security and compliance measures across the development and deployment lifecycle.
5 Conclusion and Future Directions
- GPT and Llama account for more than half of the surveyed papers, while only five discuss security and two incorporate security mechanisms.Only one paper implements countermeasures; six implement reliability-related mechanisms, while privacy-preserving principles and mechanisms receive very limited attention.Source: §5 · p. 8
- The survey identifies a gap between high functional performance and regulatory alignment, with proactive vulnerability testing, lifecycle risk governance, and training-data transparency largely absent.The authors state that isolated peak compliance scores indicate technical feasibility, while the main barrier is how research treats safety and security beyond deployability.Source: §5 · p. 8
- The authors call for interdisciplinary compliance frameworks, safe deployability for real-time mobility data, lifecycle cybersecurity, copyright adherence, and real-world stress testing.They specifically identify prompt injections, data poisoning, and cyber threats in physical deployment as risks for continuous adversarial assessment.Source: §5 · p. 8