Gartner Warns: By 2028, a Quarter of Enterprise Security Breaches Will Stem from AI Agent Abuse

Business and technology insights firm Gartner recently released the top ten strategic technology trends that enterprises and institutions need to watch in 2027. Almost all ten trends are directly or indirectly related to AI, but the focus has changed: it is no longer about what else AI can do, but about "once AI executes autonomously, who governs it, and how."

"The focus in 2026 was how to build and apply AI; the focus in 2027 shifts to how to manage AI-driven autonomous systems and redesign enterprises' technology architecture," Gao Ting, a research vice president at Gartner, told a Yicai reporter.

At the just-concluded AI parallel forum of the 2026 World Laureates Forum, global scientists also repeatedly emphasized topics such as "trustworthy AI and security governance." The industry generally believes that the AI industry will shift overall from a race in model capabilities to the building of security governance and operations systems.

The ten trends Gartner released this time can be summarized into three main lines. The first is the non-text AI frontier: large quantitative models (LQMs) are designed specifically to process and analyze quantitative data, aimed at heavily regulated industries such as finance and healthcare; voice reasoning models reason directly over audio and take action; and world models let machines understand space, time and causality.

Among these, world models are the item with the highest market heat recently, with investment in fields related to embodied intelligence, autonomous driving and spatial computing increasing substantially. "Although Google's recently announced Genie 4 demonstrates more advanced real-time world simulation capabilities, it is still in the research preview stage, and long-duration consistency and complex physical interaction remain limited," Gao Ting said. "Therefore, I believe that in the short term, commercialization opportunities are more likely to first appear in digital content and virtual scene generation; in the medium to long term, as models' physical accuracy, reliability and simulation-to-reality transfer capabilities continue to improve, industrial simulation and robotics have the opportunity to form deeper industrial value."

The second is new infrastructure for managing autonomous agents: AI context platforms, AI agent identity, AI coding agent frameworks and autonomous exposure remediation, which respectively address "whether agents understand the business, who manages permissions, who verifies code, and who fixes vulnerabilities"; the third is redefining the software economy and technological sovereignty: disposable applications make software "use once and discard," sovereign technology reserves exit channels for critical vendors, and wetware-as-a-service explores biological computing paths.

Gartner's forecast is quite cautionary: by 2028, 25% of enterprise security breach incidents will be traced back to AI agent abuse, with the root cause being that machine permissions have not been effectively managed.

Gao Ting believes priorities should be set by maturity: voice reasoning models, AI context platforms, AI coding agent frameworks and others are in a one-to-three-year implementation window, and enterprises can follow the investment principle of "creating business value in the near term, building core capabilities in the medium term, and preserving technological options in the long term."

Gartner's judgment is not isolated. Recently, the AI parallel forum of the 2026 World Laureates Forum was held in Lingang, Shanghai, under the theme "The Next Decade: The Development and Governance of Artificial Intelligence." He Jifeng, an academician of the Chinese Academy of Sciences, said in his address that he first proposed the concept of "trustworthy artificial intelligence" in 2017, with the core aim of hoping that technological development and security governance move in the same direction. He stressed that AI should never be a black box, and that by implementing correct and standardized methods, we can ensure that all parts of AI align with the human value framework. Moreover, as AI capabilities strengthen, the harm that could be caused by misuse will also grow, and safety guardrails must be reinforced in tandem.

Martin Hellman, a 2015 Turing Award laureate and professor at Stanford University, took the issue to the global level. Previously, he, together with Geoffrey Hinton, Yoshua Bengio and others, jointly signed a statement on AI risk, and he advocates listing the reduction of existential AI risk as a global priority on par with preventing nuclear war and pandemics.

Frontline industry players have also begun to act. Chen Yunji, deputy director of the Institute of Computing Technology at the Chinese Academy of Sciences, believes that current large models are based on statistical principles, and their accuracy is "always 99% and can never reach 100%," which means AI agents may cause harm. In the future, physical rules and logical rules need to be introduced to plan, constrain and drive AI to achieve 100% precision.

Embodied intelligence is accelerating into the physical world. Gao Quanzhong, co-president of Qiyuan Robot, believes that safety design must be fully considered during model training. "For example, with robots, we need to consider the people nearby and the surrounding environment, and ensure that safety is embedded from the modeling stage onward."