more

(*Japanese follows the English)

Sakana AI is launching the Security Research Group, an independent R&D team studying collective, AI-driven cyber defense and the security and safety of AI itself.

Our research starts from the idea that intelligence and resilience emerge from collectives, like a school of fish rather than a single swimmer. As a company building frontier AI, we also believe we share responsibility for the security of the society that uses it. Separate from our commercial business, the group’s mission is to serve the public interest: publishing research, reproducible benchmarks, and practical guidelines, and working with government, industry, and academia on research to improve cybersecurity.

What We Are Seeing

Frontier AI’s ability to find flaws in software is growing, and the number of disclosed flaws is climbing.1 NIST reports that CVE submissions rose 263% between 2020 and 2025, and that early 2026 ran nearly a third above the same period a year earlier.2 JVN iPedia, which provides vulnerability information in Japanese, also registered nearly twice as many entries in the latest quarter as a year earlier, most of them from the NVD.3 AI is widely suspected to play a part, though there is still more to confirm: FIRST points to AI-assisted discovery alongside administrative factors, and NIST staff say AI-based discovery tools have added to their workload.4 Japan’s government now treats the trend as a national security matter.1

The strain is already showing. NIST has stopped routine analysis of older records and now prioritizes flaws known to be exploited.5 Japan’s financial regulators have asked institutions to treat the response as a management issue, with top management directly involved.6 Flaws are being found at machine speed and fixed at organizational speed, and we suspect the people in between are the ones bearing the cost.

The other side is changing too. Attackers are using AI agents, running intrusions against many targets at once.7 And AI systems are themselves becoming targets: NICT recorded more probing aimed at AI service endpoints in the first quarter of 2026.8

Why Faster Response Is Not Enough

Faster triage and patching are necessary, but patching systems already in operation will not be enough to bring overall risk down. CISA’s review of two fiscal years finds that many attackers exploit well-known, straightforward weaknesses, with input validation and memory safety flaws the most dependable ways in, and it calls for eliminating persistent, preventable weaknesses.9 NICT’s recent alert on attacks against the servers behind mobile apps names the same kinds of weakness, such as missing authorization checks and unvalidated input.10 The FBI’s account of the AI-agent intrusions lists SSRF, IDOR, missing API authentication and injection.7 Even if AI makes finding these flaws faster, if the same weaknesses keep being introduced into software and architecture, the work is likely to keep growing.

That is why we believe the lasting answer lies upstream, in the development process of systems, AI included. Instead of chasing flaws after release, the aim is to keep them from arising at the design and build stage, or to remove them there. Japan’s government-wide package points the same way, asking software vendors to apply secure-by-design principles and to use advanced AI across development to find and fix vulnerabilities early.1

Four layers of defense, from upstream to downstream: the quality of each product, the design and configuration of the system, operation and patching, and detection and response. The Security Research Group works across all four, with its long-term weight on the upstream two.

Quality Is Where Japan Can Lead

We propose to treat this as a question of quality. Finding and fixing vulnerabilities may be seen as a special task that security specialists carry out separately, after the product is built. We would rather see it as part of the quality a product is expected to have, and raise that quality at the source. Japan may be well placed to lead on this. Japanese manufacturing built its worldwide reputation on quality, and that tradition is part of how Japan’s products have long been valued. Rethinking quality through the security of software is a challenge, and also a chance to raise Japan’s profile.

AI can help here as well. The same capability that finds flaws in finished systems can be turned into a way of checking quality at every step of design, building and operation, and we will use it to the fullest on the defending side. If the expertise of excellent human reviewers can be built into AI agents, quality could be assured at a scale and depth that were hard to reach before, and the flaws found today could teach us not to build the same flaws tomorrow. That gives defenders reason for hope.

Security Is Must-Be Quality

The Kano model, widely used in quality management, sorts quality into three kinds by how they relate to user satisfaction: attractive quality, which delights when present; one-dimensional quality, which satisfies in step with how well it is met; and must-be quality. Must-be quality goes unnoticed when it is present and causes strong dissatisfaction only when it is missing. Safety is a classic example, and security is another; because they are taken for granted and so rarely asked for explicitly, both are easily left out of the definition of quality altogether. That is why they have to be treated as quality requirements from the start and built in, rather than inspected for at the end. Keeping quality means working to never introduce the gaps that lead to incidents, in the software we run and in AI systems themselves.

Safety and security are must-be qualities in the Kano model: meeting them never lifts users above neutral, but falling short causes strong dissatisfaction. Conceptual sketch after Kano et al. (1984).

Japan’s Constraints

Responding to AI-era vulnerabilities is hard in Japan, most likely for reasons like those set out in public documents, including the request from the Financial Services Agency and the Bank of Japan.11

  • Outsourced operations: Many organizations entrust system maintenance to vendors, so whether patching is covered by the maintenance contract, and who is responsible for what, needs to be clear in advance.
  • Technical debt: Legacy systems and incomplete asset inventories can make it hard to tell which flaws apply, and harder still to fix them.
  • Limited staff: Substantially expanding IT staff in the short term is not realistic, so the expertise that exists has to be focused where it matters most.

Under these conditions, simply adding people will not keep up. We see these difficulties as structural, or design, problems to solve. The fewer flaws a product has, the less vendor coordination, patch windows and emergency staffing it needs. Keeping must-be quality through incident response after a product is built tends to be costly. If the investment goes into design and build instead, the total cost of must-be quality could fall, and the savings could go into one-dimensional and attractive quality, raising customer satisfaction. That is why we place our long-term weight upstream. A frontier model alone is not enough for this, though: without scaffolding adapted to each environment and a view of the whole system, it can produce false positives and unsafe fixes. Change does not happen overnight, which is why we want to move forward not as one organization, but together with people in Japan and around the world.

Our Research Areas

Our research has two main themes: using AI for collective defense, and the security and safety of AI itself. Drawing on secure-by-design and quality-management practice, we have the following lines of research in view:

  • Prevention by design: How can the qualities a system must keep be defined before it is built and checked for as long as it runs? How can whole classes of flaws be made unreachable? And when a fix is still needed, how can an automated fix be checked against criteria the automation itself cannot rewrite, and undone quickly if something goes wrong?
  • Collective defense with AI agents: Attackers are already using AI agents against many targets at once,7 while defense is still largely organized one organization at a time. In Japan, analysis is already shared across organizations through the efforts of the people involved, and we want to explore whether it can run at machine speed and at national scale, with defensive agents distributed across organizations that stay in sync, so that what one of them observes feeds back into how all of them defend next. An attack seen once could then become a protection shared by all who take part. The questions here are open, including what may be shared and with whose consent, how a shared signal can be trusted across organizational and vendor boundaries, and who is accountable when agents act on it.
  • Securing AI itself, starting with the defenders’ own agents: The systems behind AI are already being probed.8 Agents that triage and patch hold privileged access, so the first two areas depend on this one. How do we keep AI systems secure against attack and misuse, and safe in what they do? We start with the defenders’ own agents: how do we limit each agent’s authority to its task, trace why it acted, and keep tools built for defense bound to defense?

Our Research Foundations and Open Questions

Cybersecurity is ultimately a co-evolutionary system, with two co-evolutions at work: attack with defense, and AI’s advances with the security and safety that make those advances sustainable. In its research on self-improving AI, Sakana AI has seen failures first-hand, such as agents finding ways around the constraints they were given, and treating these as a central technical problem, not an exception, is our stance. This makes Sakana AI’s prior work a natural starting point, and our aim is to feed what these co-evolutions reveal back into design, where it becomes prevention:

  • Adversarial co-evolution (Digital Red Queen): In the programming game Core War, we evolved LLM-written programs against their own predecessors, and they tended to grow more general against unseen human-written programs. How do we bring this safely into production blue and red-teaming operations?
  • Sample-efficient search (ShinkaEvolve): Evolves stronger code within tight evaluation budgets, which we think may suit crafting detection rules and candidate patches where real-world testing is expensive.
  • Collective orchestration (Fugu-Cyber): Coordinates specialized models, reaching 86.9% on CyberGym and 72.1% on CTI-REALM, which Sakana AI reports as comparable to leading cyber-focused frontier models.12 Benchmark results do not necessarily carry over to real enterprise environments, and measuring that gap is one of our research questions.

How We Operate

To ensure our work genuinely serves Japan’s broader security ecosystem, the Security Research Group operates under these guidelines:

  • Vendor neutrality: Security research itself is, as a rule, kept separate from product promotion. When we evaluate Sakana AI’s own systems, we use the same transparent, reproducible conditions as for any outside tool, and publish compute budgets, iteration counts, harnesses and similar details as needed.
  • Responsible disclosure: We are defining our release policy. For now, our principles are to share white papers, benchmarks, and defensive playbooks openly, withhold actionable exploit material, and report discovered vulnerabilities through proper coordination channels.
  • Connecting policy and practice: Japan has well-developed security frameworks, but turning any framework into day-to-day operational practice is hard, and we aim to build the technical bridge between the two.

From the Group’s Leaders

The Security Research Group is led by Akihiro Umegai and David Ha.


“AI is changing both sides of cybersecurity at the same time, and no single organization can keep up with that on its own. I believe security and safety are not a brake on AI’s progress but the condition for sustaining it. Through this group, I will help open up research on the next generation of cybersecurity and on the security and safety of AI itself in Japan, and share what we learn with the world.”

Akihiro Umegai, Chief of Security Research, Sakana AI

“Cybersecurity is one of the clearest places where AI is already changing what is possible, for attackers and defenders alike. Defenders need tools that improve as fast as the threats do. Security is an adversarial environment that never stops changing, and that is exactly the kind of problem our work on evolution, collective intelligence, and multi-agent systems was built to study. As an AI company with roots in Japan, we also want to help protect the society we are building for. That is why we are starting this group now, and why we are setting it up to share its work openly.”

David Ha, Co-Founder and CEO, Sakana AI

Working with Sakana AI’s Security Research Group

We welcome collaboration with policy makers, enterprises, and security researchers across Japan and globally. We are also hiring! If you want to help build resilient defenses for the AI era, explore the Member of Technical Staff (Security Research Group) position on our careers page.

Japanese

Sakana AI Launches the Security Research Group

Read in English

Sakana AI is pleased to announce the launch of the Security Research Group, an independent research and development team studying collective, AI-driven cyber defense and how to ensure the safety and security of AI systems themselves.

At Sakana AI, our research has been guided by the idea that intelligent, resilient systems emerge from collectives—not from a single fish. As a company developing frontier AI, we have also felt a responsibility to protect the safety of the society that uses it. Launching today as an organization separate from our commercial activities, the Security Research Group has a mission to serve the public interest. It will conduct research and development to improve cybersecurity in collaboration with government, industry, and academia while publishing research findings, reproducible benchmarks, and practical guidelines.

What Is Happening Now

Frontier AI is becoming increasingly capable of finding flaws in software, while the number of disclosed flaws continues to rise.13 According to the U.S. National Institute of Standards and Technology (NIST), submissions for CVEs—the identification numbers assigned to publicly disclosed vulnerabilities—were nearly 30% higher at the beginning of 2026 than during the same period a year earlier.14 At JVN iPedia, which provides vulnerability information in Japanese, the number of registrations in the most recent quarter also reached nearly twice the level recorded during the same period a year earlier, although most came from the U.S. National Vulnerability Database (NVD).15

It is not yet clear how much AI has contributed to this increase. FIRST, an international association of incident response teams, cites administrative factors such as clearing backlogs in addition to AI-assisted discovery, while NIST staff have said that AI-powered discovery tools are increasing workloads.16 The Japanese government also regards these developments as a national security issue.13

The burden is already beginning to show on the front lines. NIST has stopped routinely analyzing older records and shifted to prioritizing flaws that have been confirmed as actively exploited.17 Japanese financial regulators are also asking financial institutions to address cybersecurity as a management issue requiring direct involvement from top executives.18 Organizations are currently unable to fix flaws as quickly as they are being discovered, leaving frontline personnel to bear the brunt of the strain. Attackers have begun using AI agents to attempt intrusions against large numbers of targets simultaneously,19and AI systems themselves are increasingly becoming targets. The National Institute of Information and Communications Technology (NICT) observed an increase in traffic probing AI service access points, or endpoints, in the first quarter of 2026.20

Faster Responses Alone Are Not Enough

Against this backdrop, the need to accelerate triage—the prioritization of flaws requiring attention—and patch deployment is growing. Yet continually patching systems already in operation cannot sufficiently reduce overall risk. According to the U.S. Cybersecurity and Infrastructure Security Agency’s (CISA) review of vulnerabilities spanning two fiscal years, many attackers exploit familiar, simple weaknesses rather than sophisticated techniques. In particular, failures to validate input and memory-safety flaws provide attackers with the most reliable points of entry. CISA is calling for the elimination of these long-standing weaknesses that should be preventable.21

NICT also recently warned about attacks targeting servers behind mobile applications, citing the same kinds of weaknesses, including missing authorization checks and failures to validate input.22 In cases reported by the FBI involving intrusions carried out by AI agents, weaknesses such as SSRF, IDOR, missing API authentication, and injection were also cited as entry points.19 Even if AI increases the speed at which flaws are found, the workload will only continue to grow as long as the same weaknesses keep being built into software and system architectures.

That is precisely why we believe the fundamental solution lies upstream in the development process for systems, including AI systems. Instead of chasing flaws after release, the goal should be to avoid creating them in the first place—or eliminate them immediately—during the design and construction stages. The Japanese government’s policy also calls on software vendors to apply secure-by-design principles and use advanced AI throughout the development process to identify and fix vulnerabilities early.13

Four layers of defense extending from upstream to downstream: the quality of individual products; system design and configuration; operations and patching; and detection and response. The Security Research Group will work across all four layers while placing greater long-term emphasis on the two upstream layers.

Quality as a Japanese Strength

We propose framing this as a “quality problem.” Vulnerability discovery and remediation are often regarded as specialized tasks performed separately by security experts after a product has been completed. We believe instead that security should be treated as an integral aspect of product quality—and that this quality should be improved from its source.

Japanese manufacturing has earned the world’s trust through quality. Now that AI capabilities are improving, we can rethink the concept of quality from a security perspective. We believe this presents both a major challenge and an opportunity for Japan to strengthen its presence. AI will also support this effort. AI’s ability to find defects in completed systems can be used to verify quality at every stage—design, construction, and operation. If the expertise of outstanding reviewers can be incorporated into AI agents, it may become possible to verify quality at a scale and depth that were previously unattainable. If we can create a cycle of “learning from the defects found today so that we do not create the same defects tomorrow,” perhaps defenders, too, will begin to see hope.

Security Is “Must-Be Quality”

The Kano model, widely used in quality management, divides quality into three categories based on how it relates to user satisfaction: attractive quality, which delights users when present; one-dimensional quality, for which satisfaction increases in proportion to the degree of fulfillment; and must-be quality. Must-be quality goes unnoticed when present but causes strong dissatisfaction when absent. Safety is a typical example, and security is another. Because it is taken for granted and therefore rarely requested explicitly, it can easily be omitted from the definition of quality. That is precisely why it must not be inspected only at the end, but instead identified from the outset as a quality requirement and built into the product. Preventing opportunities for incidents from arising in the first place—both in the software being used and in the AI system itself—is essential to maintaining quality.

Safety and security constitute must-be quality in the Kano model. Even when fully satisfied, they do not raise users’ assessments above neutral, but failing to satisfy them causes dissatisfaction. Conceptual diagram based on Kano et al. (1984).

Constraints Unique to Japan

Japan also faces distinctive challenges in responding to vulnerabilities in the AI era. Requests from the Financial Services Agency and the Bank of Japan, among others, point to the following circumstances.23

  • Outsourced operations:Many organizations entrust system maintenance to vendors. It is necessary to clarify in advance whether patching falls within the scope of the contract and who is responsible for what.
  • Technical debt:Because legacy systems remain in use and assets are not fully inventoried, it is difficult to determine whether a newly discovered defect applies to an organization’s own systems, and remediation also requires considerable effort.
  • Limited personnel:It is unrealistic to substantially increase IT staffing in a short period, so existing specialists must be concentrated where they are most needed.

Under these conditions, simply adding more people will not be enough to keep pace. We believe this is a structural problem that requires reexamining the very mechanisms by which defects arise and addressing them at the design stage. If products have fewer defects from the outset, less vendor coordination, fewer patching windows, and fewer personnel for emergency response will be needed. Attempting to preserve must-be quality through incident response after a product has been completed incurs costs every time a defect is discovered. By investing in design and manufacturing to build in must-be quality, these expenditures can be reduced, freeing resources to improve one-dimensional and attractive quality and thereby increase customer satisfaction. This thinking is why we place our long-term emphasis upstream.

That said, this cannot be achieved merely by deploying frontier models. Without scaffolding suited to each environment and a perspective that encompasses the entire system, AI may produce false positives or unsafe fixes. Such a transformation will also take time. That is why we want to pursue it in collaboration with stakeholders in Japan and abroad, rather than confining the effort to a single organization.

Research Areas

The Group has two main research pillars: AI-enabled collective defense, and ensuring the safety and security of AI systems themselves. Building on expertise in secure by design and quality management, we are considering research in the following areas.

(1) AI-Enabled Collective Defense

  • Prevention by design:How should the quality required of a system be defined before construction, and how can it continue to be verified during operation? How should systems be designed to make certain types of defects structurally unusable in attacks? When fixes are nevertheless required, how can the safety of automated remediation be verified, and how can changes be promptly rolled back if problems arise?
  • Collective defense through AI agents:Attackers are already using AI agents to target large numbers of systems simultaneously.19 Meanwhile, much of defense still remains confined within individual organizations. In Japan, thanks to the efforts of those involved, analytical findings have continued to be shared across organizational boundaries. The question is how to make this initiative operate at machine speed and on a nationwide—or even global—scale. What we envision, for example, is a system in which defensive agents deployed at individual organizations coordinate with one another, so that an attack observed at one organization is immediately reflected in defenses across the entire network. If realized, information about an attack observed once could be used to protect every participating organization. However, questions remain unanswered: what should be shared and under what form of consent; how can the reliability of shared information be ensured across organizational and vendor boundaries; and who bears responsibility when agents act on that information?

(2) Ensuring the Safety and Security of AI Systems Themselves

  • Protecting AI systems themselves (starting with defensive agents):The systems supporting AI are already being probed by attackers.20 In particular, agents responsible for triage and patching have privileged access to systems. Unless the safety of these agents is ensured, the two areas above cannot function either. We will therefore begin our research into protecting AI systems from attacks and misuse with defensive agents. For example: how can each agent’s permissions be limited to the scope of its assigned duties; how can the basis for its actions be made verifiable after the fact; and how can tools developed for defense be prevented from being repurposed?

Research Foundations and Open Questions

Offense and defense in cybersecurity have a “coevolutionary” relationship: when one evolves, the other evolves as well. Coevolution is also necessary between AI and its safety. Unless security and safety advance alongside AI, progress in AI itself will not be sustainable.

In our research on self-improving AI, Sakana AI has encountered failures such as agents finding loopholes in the constraints imposed on them. Rather than dismissing such failures as exceptions, we treat them as central technical challenges. Evolution and coevolution are also themes that Sakana AI has placed at the heart of its research. We will apply the methods developed through that work, incorporating weaknesses revealed during the coevolutionary process into system design to help prevent defects. The following three research projects are our starting point.

  • Adversarial coevolution (Digital Red Queen):In the programming game Core War, we evolved LLM-written programs by having them compete against their past versions. Over time, they showed a tendency to become broadly stronger even against previously unseen programs created by humans. The next challenge is determining how to safely introduce this approach into real-world blue-team and red-team operations.
  • Sample-efficient search (ShinkaEvolve):A method for evolving stronger code under a constrained evaluation budget. It may be well suited to creating detection rules and proposed fixes, where testing in real-world environments is costly.
  • Collective orchestration (Fugu-Cyber):Fugu-Cyber, which coordinates specialized models, demonstrated performance comparable to leading frontier models specialized in cybersecurity.24 However, benchmark results do not necessarily translate directly to actual enterprise environments, and measuring that gap is itself a research challenge.

Operating Principles

To ensure that our work genuinely benefits Japan’s entire security ecosystem, the Security Research Group will operate according to the following principles.

  • Vendor neutrality:As a rule, the group’s security research will be conducted independently of product promotion. When evaluating Sakana AI’s own systems, we will do so under the same transparent and reproducible conditions used for external tools, disclosing the compute budget, number of iterations, harnesses, and other details as appropriate.
  • Responsible disclosure:We will formulate a policy for publishing research results in the future. For the time being, as a general rule, we will publish white papers, benchmarks, and defensive playbooks while withholding materials for exploits that could be misused. Any vulnerabilities we discover will be reported through the appropriate coordinating organizations.
  • Bridging policy and practice:Although Japan has well-developed security frameworks, every organization faces challenges in translating those frameworks into day-to-day operations. We aim to serve as a technical bridge between policy and operational practice.

From the Group Leaders

The Security Research Group is led by Akihiro Umetani and David Ha.


“AI is transforming both the offensive and defensive sides of cybersecurity at the same time, and no single organization can keep pace on its own. We believe that safety and security are not impediments to progress in AI, but prerequisites for sustaining that progress over the long term. Through this group, we will pioneer research in Japan on next-generation cybersecurity and on the security and safety of AI itself, and share the resulting insights with the world.”

Akihiro Umetani, Chief of Security Research, Sakana AI

“Cybersecurity is one of the fields where AI is most visibly transforming what is possible, on both the offensive and defensive sides. Defenders need tools that evolve as quickly as the threats. Security is a constantly changing adversarial environment—exactly the kind of problem we have been tackling through our research into evolution, collective intelligence, and multi-agent systems. As an AI company rooted in Japan, we also want to help protect the society we are working to build. That is why we are launching this group now and will openly share its成果.”

David Ha, Co-founder and CEO, Sakana AI

Collaborating with the Group and Joining Our Team

We welcome collaboration with policymakers, companies, and security researchers in Japan and around the world. We are also hiring. If you would like to help us build resilient defenses for the AI era, please see theCareers page“Member of Technical Staff (Security Research Group)” position.

Footnotes

  1. Project YATA-Shield is the government-wide package issued jointly on May 18, 2026 by the National Security Secretariat, the National Cybersecurity Office and twelve other ministries and agencies. It is premised on the rapid improvement of frontier AI models at finding and fixing vulnerabilities, and states that misuse of AI by attackers is raising the speed and scale of cyberattacks. It asks critical-infrastructure operators to strengthen measures on the premise that vulnerabilities will be found and fixed faster and in larger numbers, and alerts software vendors to apply secure-by-design principles and use advanced AI across the software development lifecycle to find and fix vulnerabilities early. Sources: package and summary. ↩

  2. NIST cited a 263% rise in CVE submissions between 2020 and 2025, and said that submissions in the first three months of 2026 were nearly one-third higher than a year earlier. Source: NIST. ↩

  3. JVN iPedia registered 20,073 entries in the third quarter of 2026, against 10,869 a year earlier; 19,913 of them were collected from the U.S. National Vulnerability Database (NVD), 147 from JVN and 13 from Japanese product developers. Counts follow the date of registration, which follows the timing of publication in NVD. Source: IPA, JVN iPedia registration status, third quarter of 2026. ↩

  4. FIRST’s mid-year update (June 15, 2026) attributes the rise to AI-assisted discovery together with administrative factors such as the clearing of CNA backlogs and the expansion of GitHub Security Advisories. NIST staff said in April 2026 that increased submissions driven by LLM-based discovery tools have added to the NVD’s workload. Sources: APNIC, summarizing FIRST; Infosecurity Magazine. ↩

  5. The NVD has stopped routine enrichment of unenriched records published before March 1, 2026, and prioritizes records on CISA’s Known Exploited Vulnerabilities catalog, software used within the federal government, and critical software as defined by Executive Order 14028. Sources: NIST; Infosecurity Magazine. ↩

  6. On May 22, 2026, the Financial Services Agency and the Bank of Japan asked financial institutions and related entities to treat the response to frontier AI as a management issue with the direct involvement of top management, and to make patching risk-based. Sources: Financial Services Agency and Bank of Japan. ↩

  7. FBI FLASH-20261113-001, “Threat Overview of ARTEX Autonomous Penetration Testing Framework,” November 13, 2026 (TLP:CLEAR). It describes ARTEX as an open-source, LLM-driven, multi-agent platform that threat actors have weaponized to automate multiple phases of intrusions at machine speed, in use since at least August 2026, and says it lets actors with minimal expertise run complex multi-target operations simultaneously. The figure of intrusions reaching privileged access in under three hours is cited in the FLASH to Microsoft Threat Intelligence. The initial-access weaknesses it lists include SSRF, IDOR, missing authentication on web APIs, command injection, SQL injection, and hardcoded credentials; the FLASH names no specific CVE identifier. Source: FBI IC3. ↩

  8. NICTER recorded a rise in probing for AI service endpoints in the first quarter of 2026. Darknet observation records unsolicited traffic and does not count attacks. Source: NICT, NICTER observation statistics, January to March 2026. ↩

  9. CISA’s Vulnerability Review, published August 26, 2026, covers fiscal years 2024 and 2025. It finds that many attackers target well-known, straightforward weaknesses rather than sophisticated techniques, that input validation and memory safety flaws are the most dependable entry points, and recommends eliminating persistent, preventable weaknesses and adopting Secure by Design principles. Sources: CISA announcement; full review. ↩

  10. The alert comes from NICT’s LETTICE team, a joint analysis team of volunteers from organizations participating in NICT’s Cybersecurity Nexus (CYNEX). It identifies the A2B (App-to-Backend) attack as one cause of a recent series of data leaks in Japan: attackers analyze apps published in Japan to learn the backend servers and APIs behind them, then attack each backend. It names four weakness classes as mainly targeted: SSRF, remote code execution, SQL injection and IDOR. Source: NICT CYNEX, alert of October 13, 2026 (Rev. 1.1). ↩

  11. All three conditions appear in the request from the Financial Services Agency and the Bank of Japan cited in footnote 6. It observes that many financial institutions entrust system maintenance to vendors and asks them to confirm that vendors can secure the resources to meet service-level agreements when patching is needed at several institutions at once. It asks institutions to clear technical debt in priority systems so that the targets of a patch can be identified immediately. And it states that substantially expanding existing IT staff in the short term is not realistic. The request addresses the financial sector. ↩

  12. Fugu-Cyber reached 86.9% on CyberGym and 72.1% on CTI-REALM, as reported by Sakana AI; the page gives no baseline scores or evaluation setup. Source: Sakana AI, “Introducing Fugu-Cyber”. ↩

  13. The Japanese government’s Project YATA-Shield package also reflects the recognition that frontier AI is becoming increasingly capable of finding vulnerabilities. Jointly released on May 18, 2026, by the National Security Secretariat, the National Cybersecurity Office, and twelve other ministries and agencies, the package is premised on the rapid improvement of frontier AI models’ ability to discover and fix vulnerabilities. It states that attackers’ misuse of AI is increasing the speed and scale of cyberattacks. It therefore asks critical-infrastructure operators to strengthen their measures on the assumption that vulnerabilities will be discovered and fixed faster and in greater numbers. It also cautions software vendors to apply secure-by-design principles and use advanced AI throughout the software development lifecycle to discover and fix vulnerabilities early. Sources:package、summary ↩

  14. NIST cited a 263% increase in CVE submissions between 2020 and 2025 and said submissions during the first three months of 2026 were nearly one-third higher than in the same period a year earlier. Source:NIST ↩

  15. JVN iPedia recorded 20,073 entries in the third quarter of 2026, compared with 10,869 in the same period a year earlier. Of these, 19,913 were collected from the U.S. National Vulnerability Database (NVD), 147 from JVN, and 13 from domestic product developers. The figures are based on registration dates, and the timing of registration is affected by when records are published in the NVD. Source:IPA, “JVN iPedia Registration Status (Third Quarter of 2026)” ↩

  16. FIRST’s interim update of June 15, 2026, cited administrative factors behind the increase—including the clearing of CNA backlogs and the expansion of GitHub Security Advisories—in addition to AI-assisted discovery. In April 2026, NIST staff said that increased submissions from LLM-based discovery tools were adding to the NVD’s workload. Source:APNIC (summary of FIRST)、Infosecurity Magazine ↩

  17. The NVD has stopped routinely enriching unenriched records published before March 1, 2026, and is prioritizing records listed in CISA’s Known Exploited Vulnerabilities Catalog (KEV), software used within the federal government, and critical software designated under Executive Order 14028. Source:NIST、Infosecurity Magazine ↩

  18. On May 22, 2026, the Financial Services Agency and the Bank of Japan called on financial institutions and others to treat frontier AI preparedness as a management issue requiring the direct involvement of senior leadership and to adopt a risk-based approach to patching. Sources:Financial Services Agency、Bank of Japan ↩

  19. FBI FLASH-20261113-001, “Threat Overview of ARTEX Autonomous Penetration Testing Framework,” November 13, 2026 (TLP:CLEAR). It describes ARTEX as an open-source, LLM-driven, multi-agent platform. It states that attackers have been misusing it since at least August 2026 to automate multiple stages of intrusions at machine speed, enabling even less-skilled attackers to conduct complex, multi-target operations simultaneously. The figure of less than three hours for an intrusion to achieve privileged access is cited in the FLASH from Microsoft Threat Intelligence materials. The initial-access weaknesses listed in the FLASH include SSRF, IDOR, missing Web API authentication, command injection, SQL injection, and hard-coded credentials; it does not identify individual CVE numbers. Source:FBI IC3 ↩

  20. In the first quarter of 2026, NICTER observed an increase in scanning targeting AI service endpoints. Darknet monitoring records unsolicited communications; it does not count the number of attacks. Source:NICT, NICTER Observation Statistics (January–March 2026) ↩

  21. CISA’s vulnerability review, published on August 26, 2026, covers fiscal years 2024 and 2025. It notes that many attackers target simple, well-known weaknesses rather than using sophisticated techniques, and that inadequate input validation and memory-safety flaws provide the most reliable points of entry. It recommends eliminating persistent, preventable weaknesses and adopting secure-by-design principles. Sources:CISA announcement、full review ↩

  22. This advisory was issued by NICT’s LETTICE team. LETTICE is a joint analysis team composed of volunteers from organizations participating in NICT’s Cybersecurity Nexus (CYNEX). It identifies A2B (App-to-Backend) attacks as one cause of a series of data breaches in Japan. Attackers analyze publicly available apps in Japan to identify the backend servers and APIs behind them, then attack each backend. It identifies four main types of targeted weakness: SSRF, remote code execution, SQL injection, and IDOR. Source:NICT CYNEX advisory dated October 13, 2026 (Rev. 1.1) ↩

  23. All three conditions are set out in the request from Japan’s Financial Services Agency and the Bank of Japan cited in footnote 18. The request notes that many financial institutions outsource system maintenance to vendors and asks them to verify whether vendors can secure sufficient resources to meet service-level agreements even when multiple financial institutions need patches applied simultaneously. It also calls for addressing technical debt in high-priority systems so that patching targets can be identified immediately, while noting that it is unrealistic to substantially increase existing IT staffing within a short period. The request applies to the financial sector. ↩

  24. Fugu-Cyber achieved 86.9% on CyberGym and 72.1% on CTI-REALM. Source:Sakana AI「Introducing Fugu-Cyber」 ↩