{{item.videoDuration}}
{{item.title}}
{{item.text}}
{{item.videoDuration}}
{{item.text}}
AI is shifting from an entity that "thinks" to one that "acts." The rise of the "AI agent"—which plans tasks towards a goal, connects to external systems, and executes based on its own judgement—delivers significant value to enterprises, but it also creates cyber risks that are qualitatively different from earlier uses of AI.
So what specific risks arise when AI agents become embedded in enterprise systems and business processes? And how far have Japanese enterprises progressed in preparing for them? This report clarifies the risk structure unique to AI agents and presents a practical governance framework that Japanese enterprises should begin implementing now.
How AI agents change the nature of risk
AI agents reason, draw on memory, select tools and act on external systems. The issue is no longer confined to the quality of AI output. With what privileges, referencing what information, through what judgements, and on which systems does the AI act? What is now in question is the governance of behaviour itself.
Even well-known threats such as prompt injection and data poisoning can, in an AI agent, propagate through cognition, memory and tool execution, ultimately surfacing as real damage to business systems. This report treats the AI agent not as a mere extension of conventional AI but as a "system that acts autonomously," and examines—drawing on experimental results—how this characteristic expands and transforms risk.
Rising expectations, and lagging control readiness of Japanese enterprises
PwC Japan Group's Generative AI Survey 2026 Spring: A Six-Country Comparison (covering respondents at the section-manager level or above at companies with revenue of ¥50 billion or more across six countries—932 respondents in Japan and 2,112 across the other five countries) shows that AI agent adoption has already entered the execution stage. The share of companies that "understand AI agents and have adopted them or are adopting them" stands at 59% in the US, 55% in the UK, 48% in China, 37% in Germany and 36% in Korea. Japanese enterprises, by contrast, lag behind at just 33%. More concerning still is the delay in control foundations: only 23% of Japanese companies have established a central organisation for AI governance (composed of the first and second lines of defence)—far below the US (63%) and the UK (54%), and the lowest figure among the six countries. The essence of this gap is not simply a shortage of budget. It reflects a compounding of two factors: a shortage of talent experienced in designing controls for the autonomous action chains of AI agents, and a structural lag in governance foundations.
Understanding AI agent risk as a structure
To address these challenges, this report systematically organises AI agent risk using the following framework:
- Five levels of autonomy (L1–L5): As AI moves up the levels from "assistive" to "fully autonomous," risk shifts from a matter of output quality to a matter of governing behaviour.
- Five functional domains: Risk is captured as a chain across five domains—Governance & Identity, Supervision & Orchestration, Action & Tools, State & Memory, and Cognition & Reasoning.
- Mapping of the threat taxonomy (T1–T17): OWASP's taxonomy1 is mapped onto a framework of autonomy level x functional domain, making the overall picture of risk visible.
Beyond this theoretical organisation, the report also demonstrates the reality and impact of these risks—both quantitatively and qualitatively—through proof-of-concept experiments based on concrete attack scenarios.
What Japanese enterprises should begin now
Centred on the AI Control Assurance Level (AI-CAL), this report systematically organises preventive, detective and responsive controls, and puts forward concrete, prioritised actions for Japanese enterprises to take. It presents a roadmap tied to each adoption stage, clarifying—from a practical standpoint—where to begin.
The adoption of AI agents is a dividing line that determines whether the value of generative AI is realised. According to the Generative AI Survey 2026 Spring: A Six-Country Comparison, among companies reporting that the effect of generative AI "far exceeded expectations," AI agent adoption rates reach 67–83%, whereas among companies whose results fell "short of expectations," the rate is only 17–31%. The first step towards using AI agents safely and effectively is to understand—and begin implementing—control designs commensurate with their autonomy and connectivity.
The age of autonomous AI: Cyber risks of AI agents and practical governance
{{filterContent.facetedTitle}}
{{contentList.dataService.numberHits}} {{contentList.dataService.numberHits == 1 ? 'result' : 'results'}}
{{contentList.loadingText}}
{{contentList.loadingText}}
