Business-adoption surveys and security experiments answer different questions. The first describes respondents’ reported use; the second tests particular systems and attacks in defined environments.
AI agent risk evaluation and technical threats
The integration of agentic AI introduces specific technical vulnerabilities. According to the National Cyber Security Centre, as Large Language Models (LLMs) increasingly share data with third-party applications, the risk of malicious prompt injection grows. These attacks, alongside data poisoning, can be difficult to detect. Furthermore, researchers at NIST highlight that agent hijacking remains a persistent challenge as these systems evolve, requiring robust evaluation frameworks to manage potential security gaps.
Barriers to Adoption and System Oversight
The UK adoption study’s fieldwork ran from February to May 2025. Among businesses already using AI, it reported 7% using agentic AI and 67% applying significant input or checking to AI outputs. The oversight figure concerns AI use generally, rather than agent security in particular. These self-reports are not evidence that checking prevented attacks or that the same adoption rates apply today.
Industry Perspectives on Agent Security
TechCrunch reports Microsoft’s proposal for separate monitoring and control of enterprise agents. The company’s commercial role should be distinguished from independent evaluation of the product. NIST’s agent-hijacking work similarly emphasises testing the actual attack and task, rather than inferring safety from interest in the technology or a single overall statistic.
Sources and further reading
Thinking about the security of AI systems | National Cyber Security Centre
https://www.ncsc.gov.uk/blog-post/thinking-about-security-ai-systems
AI Adoption Research - GOV.UK
https://www.gov.uk/government/publications/ai-adoption-research/ai-adoption-research
Microsoft says AI agents are like insider threats as 'frontier models can't be trusted' | TechCrunch
https://techcrunch.com/2026/10/19/microsoft-says-ai-agents-are-like-insider-threats-as-frontier-models-cant-be-trusted/
Technical Blog: Strengthening AI Agent Hijacking Evaluations | NIST
https://www.nist.gov/news-events/news/2025/01/technical-blog-strengthening-ai-agent-hijacking-evaluations