A government-commissioned review by Lancaster University researchers has identified significant blind spots in AI security research, mapping gaps across five critical areas that demand urgent attention. The team, invited by the Department for Science, Innovation and Technology (DSIT), analyzed more than 9,100 peer-reviewed sources published since 2020 to pinpoint where research, policy, and investment are falling short.
The review, published on July 14, 2026, used data-driven techniques to extract key themes and quantify the gaps. The findings reveal that AI security research remains narrowly focused on the models themselves, often overlooking the full lifecycle of AI systems and the infrastructure they depend on.
Five key areas needing further research
The analysis identified five domains where AI security knowledge is currently thin.
Data integrity: Research must develop formal methods to guarantee or verify the safety, authenticity, and security of data used by AI systems, including training data.
Security of AI system infrastructure: More attention is needed on the infrastructure underpinning generative or agentic AI systems to understand how large language models alter what it means to be secure by design.
Securing against end-user risks: There is little research on how end users can pose a risk to, or be at risk from, AI systems when they rely on hallucinated or malicious AI outputs.
Data and model provenance: Understanding how third-party AI models are developed and trained, and their associated security implications, is an area with significant gaps.
Model disposal: Safe decommissioning of AI systems at the end of their lifespan has received almost no research attention.
For professionals working in AI for Science & Research, these gaps represent a direct call to action. The review underscores that securing AI goes far beyond algorithmic robustness, touching every stage from data collection to retirement.
Agentic AI: an emerging threat
The review highlights agentic AI systems as a particularly concerning blind spot. These models, which can act autonomously on behalf of users, are expected to become more prevalent, yet security research has not kept pace. The authors anticipate that research in this area will grow as the technology spreads.
Lead author Dr. Edward Austin, a research fellow in Lancaster University's School of Computing and Communications, said, "The review findings highlight how the cybersecurity of AI goes beyond the models themselves, and that more research is needed to understand and secure these systems across their life cycle. With AI adoption across society increasing, this is crucial for the safe use of this technology."
Professor Nick Race, associate dean for research in Lancaster University's Faculty of Science and Technology, added, "We were delighted to be invited by DSIT to lead a review into the AI security research landscape, supporting the government in its work to enhance the cyber resilience of the UK."
The full thematic review is available on the UK government's website: Thematic review and gap analysis on AI security.
Why this matters for science and research professionals
For researchers and scientists integrating AI into workflows, the identified gaps carry practical consequences. Untested assumptions about data integrity or model provenance can compromise experimental reproducibility. Lack of infrastructure security research leaves systems that handle sensitive data vulnerable. The review's emphasis on the full lifecycle also means that disposal-often an afterthought in academic settings-must become part of security planning. Anyone building or deploying AI tools in research environments should study this gap analysis to align their work with emerging security priorities. For cybersecurity specialists, the findings map directly to skills covered in the AI Learning Path for Cybersecurity Analysts, where securing AI systems across their lifecycle is a core competency.
Your membership also unlocks:
