Among all the advances in artificial intelligence, “frontier AI” is now attracting the greatest attention. These cutting-edge models combine advanced reasoning, understanding across multiple forms of information, and the ability to carry out assigned goals autonomously. They could rapidly lift industrial productivity, but they also carry serious risks, including misuse for cyberattacks and weapons of mass destruction.
Benefits and threats are two sides of the same technology. This “dual-use” nature is essential to understanding frontier AI. This article outlines its definition and technical characteristics, its implementation in industry, and the rapidly evolving international governance landscape.
What Is Frontier AI?
“Frontier AI” is not the name of a particular product. It is a relative term for the models with the most advanced and powerful capabilities at a given point in time. It is easiest to understand as a dynamic concept whose contents change as technology progresses. Whereas conventional AI was designed for narrow, specialized tasks such as image classification and fraud detection, frontier AI operates as a general-purpose foundation model.
A major technical breakthrough is “multimodal integration,” which handles text, images, audio, and video simultaneously within one architecture instead of processing each separately.
Frontier AI has also demonstrated a phenomenon known as “emergent capabilities.” When a model exceeds a certain scale, new reasoning abilities that even its developers did not anticipate may suddenly appear. This unpredictability is both a strength and the source of the greatest challenge: difficulty of control.
From Chatbots to Autonomous Agents
When ChatGPT was released in November 2022, its role remained largely within the bounds of a “chatbot” that returned answers to prompts. It was mainly an assistive tool: users might have it generate a short piece of code, then paste that code manually.
From 2025 to 2026, however, attention shifted dramatically toward “agentic AI.”
Today's leading models function as “autonomous agents.” They can plan across multiple steps, call external tools and APIs, and carry out hours of work with minimal human involvement.
The progress is also visible in the numbers. According to an evaluation in Stanford University's AI Index Report 2026, scores on a leading software-development benchmark rose in just one year from 60 percent to nearly 100 percent.
The success rate for complex development tasks lasting more than an hour, reported as below 5 percent in late 2023, has also improved substantially. From a market perspective, repeated forecasts suggest that the field of AI agents will expand severalfold over the next few years.
If agents reach the stage of “closing the loop” by designing and improving models themselves, some observers say that self-improvement beyond human design could begin in the near future.
The U.S.–China Performance Gap Has Almost Disappeared
The development of frontier AI is also the central arena in the contest for technological supremacy. According to the AI Index Report 2026, the performance gap between leading U.S. and Chinese models has effectively disappeared. China's DeepSeek-R1 briefly matched the leading U.S. model in February 2025, and the two countries have traded the top position several times since.
As of March 2026, a leading model from the U.S. company Anthropic held a margin of only 2.7 percent in an exceptionally close race. The United States retains an advantage in the number of top-tier models and influential patents, while industry adoption has reached 88 percent.
China, meanwhile, leads the world in numbers of papers, citations, patent applications, and industrial-robot deployments, establishing a firm position through the scale of implementation. A development race once dominated by the United States has become a multipolar struggle for leadership.
Transforming Science and Medicine
The most symbolic scientific achievement is AlphaFold 3, announced by Google DeepMind and Isomorphic Labs in May 2024. The previous AlphaFold 2 dramatically improved protein-structure prediction with its Evoformer architecture. AlphaFold 3 adopted a next-generation architecture called Pairformer, enabling highly accurate prediction of interactions across biomolecules, including not only proteins but also DNA, RNA, and small molecules used in medicines.
The 2024 Nobel Prize in Chemistry awarded to the developers of AlphaFold symbolizes an era in which AI leads the frontier of basic science. In assessments of specialist knowledge, frontier AI has reportedly already surpassed PhD-level experts in accuracy in fields such as chemistry and biology, while also demonstrating strong capacity to assist with laboratory troubleshooting.
Limits remain, however. Accuracy still declines for unknown molecules unlike the training data. Adoption is also advancing across healthcare, from disease detection and drug discovery to personalized medicine, and systems have appeared that take on enormous administrative workloads such as processing referrals and managing waiting lists.
At the same time, medical data fragmented by country, difficulties with cross-border access, and safe integration and validation in clinical settings remain major hurdles.
“AI versus AI” in Cyberspace
Cybersecurity is where the dual-use character of frontier AI appears most starkly. Attackers can use its autonomy and reasoning ability to run an end-to-end process, from finding unknown vulnerabilities and generating exploit code to automating phishing on a massive scale.
AI's cybersecurity capabilities reportedly rose rapidly from the level of an inexperienced apprentice in 2023 to a level capable of tasks comparable to those of a skilled practitioner in 2025.
Japan has responded at the national level. On May 22, 2026, the Financial Services Agency and the Bank of Japan jointly requested “Short-Term Responses by Financial Institutions and Others in Light of Changes in Threats from Frontier AI.”
Recognizing the seriousness of automated vulnerability discovery and increasingly sophisticated attack methods, the unusual request called for organizations to treat the issue as a management concern rather than one limited to IT departments and to promptly inspect and strengthen asset management, patching, monitoring, and response systems. Defenders have also begun responding with frontier AI.
In July 2026, NTT DATA and NTT DOCOMO BUSINESS announced that on July 31 they would launch a cyber-risk response service incorporating multiple frontier AI models. It provides continuous support from threat and vulnerability discovery through impact assessment, remediation, and ongoing operations. AI presents possible responses, and specialist teams validate them and determine priorities, enabling organizations to implement high-priority measures with limited staff.
BroadBand Security, which specializes in information security, is also strengthening monitoring, detection, and response services aligned with the same request. The dual-use structure peculiar to cyberspace—countering offensive misuse with defensive use—emerges clearly here.
Progress in International Cooperation and a Rift in Paris
Countries have moved at unprecedented speed to build governance for powerful frontier AI. The starting point was the G7 Hiroshima AI Process agreed in autumn 2023. It established international guiding principles and a code of conduct for developing organizations, comprehensively calling for risk assessment and mitigation as well as watermarking of AI-generated material. At the Seoul Summit in South Korea in May 2024, 16 leading AI companies subsequently announced the Frontier AI Safety Commitments, promising measures including predeployment evaluation of models carrying severe risks.
The direction changed at the Paris AI Action Summit in France in February 2025. The emphasis shifted sharply from safety and risk management toward innovation and economic opportunity. As countries attracted enormous investment, proposals for specific risk thresholds promised in Seoul and mechanisms to track corporate accountability were set aside.
The United States and United Kingdom declined to sign the French-led joint statement, bringing differences between countries—whether to emphasize safety or prioritize industrial competitiveness—into the open as a clear division.
As cooperation among countries has stalled, model developers have rushed to publish voluntary safety policies. An expanding movement establishes “capability thresholds” and “risk thresholds” and defines in advance the measures to take as catastrophic risk approaches.
The Regulatory Frontier: California and Japan
With a comprehensive international treaty difficult to achieve, frameworks with practical binding force are taking shape through national and state laws. California is at the forefront.
After Governor Newsom vetoed SB 1047, a 2024 safety bill criticized for being too broad, lawmakers reworked the balance between transparency and effectiveness. SB 53, the Frontier Artificial Intelligence Transparency Act, was enacted in September 2025 and took effect in January 2026.
The law defines catastrophic risk to include more than 50 deaths or injuries or more than $1 billion in damage from a single incident, as well as expert assistance with the production of advanced CBRN weapons. It requires annual publication of risk-assessment frameworks, transparency reports before model deployment, and whistleblower protection, and imposes civil penalties of up to $1 million per violation.
Rather than imposing technical restrictions such as a kill switch, this approach seeks to reduce risk through transparency and corporate governance and is expected to become a model for future legislation. Japan is also building institutions. In February 2024, the AI Safety Institute (AISI) was established under the Information-technology Promotion Agency to develop methods and standards for safety evaluation.
It published guides to evaluation perspectives and red-teaming methods in September 2024 and signed a memorandum of cooperation with Anthropic in October 2025 to share evaluation best practices. By hosting an international forum on trustworthy AI in Hiroshima in January 2026, it has advanced domestic standards and international cooperation in tandem.
Questions for the Future
Frontier AI is becoming embedded in society as “intellectual infrastructure” that solves difficult problems in basic science and pushes the limits of healthcare and software development. Looking ahead reveals three issues.
First, “recursive self-improvement,” in which agents design and improve themselves, is becoming increasingly realistic. It could create unforeseen threats that predeployment risk assessments cannot fully capture.
Second, automating defense is no longer optional but essential. As attacks become automated and faster, defenders have no choice but to move toward resilient systems that incorporate frontier AI.
Third is the shift in the center of regulatory gravity from “limits on technical capability” to “institutionalized transparency and accountability.”
To maintain public trust beyond the division exposed in Paris, the key may lie less in constraints such as mandatory kill switches than in transparent corporate governance through mechanisms such as transparency reports and whistleblower protection.
To receive the benefits while avoiding catastrophic risks, countries, industry, and academia must now work together on dynamic, resilient governance that continually updates both technical safeguards and institutional transparency.
