top of page

Cyber Security Report -May 2026

  • Jun 18
  • 13 min read

Research by OSINT Monitoring Team, NTT Security Japan K.K


Table Contents


This report selects and summarizes three particularly important topics from various incidents and events related to information security that occurred during May 2026, as well as the changes in the surrounding environment. The main points of each topic are as follows.


Chapter 1: 'Who owns powerful AI? Democratizing AI Design Demonstrated by OpenMythos'


A 22-year-old individual developer attempted to reconstruct Anthropic's private AI model "Claude Mythos" based on a third-party academic paper. This has been released as the open-source project "OpenMythos," attracting attention from researchers and others.


OpenMythos has no official relationship with Anthropic, and there is no guarantee that the content accurately reflects the structure of Claude Mythos. On the other hand, Forbes reports that the fact that the internal design of AI models can be reconstructed from public information is significant in itself.


The emergence of OpenMythos has shown that the fundamental question of "who can have powerful AI" may be changing. Going forward, in a society where the use of powerful AI is expanding, it will be necessary to steer the company and organizations to update their defense postures.


Chapter 2: 'Proximity Attack on Taiwan High-Speed Rail'


In April 2026, a Taiwanese university student used commercially available software-defined radio (SDR) equipment to compromise the communication system of Taiwan High Speed Rail and transmitted a fake emergency alert signal, causing train operations to be halted.


By using software-defined radio via commonly available devices, it exploited operational management flaws such as prolonged inability to update authentication information and communication settings, enabling the transmission of signals disguised as legitimate control signals even without advanced attack techniques.


Similar wireless communication is widely used, and system breaches from radio should be considered to be expected not only in the railway sector but also in critical infrastructure systems.


Chapter 3: 'Structural shift of Bug Bounties Driven by AI'


Companies like Google in the US are reviewing bug bounties in response to changes in AI-driven vulnerability discovery processes.


While AI has enabled more efficient investigations and expanded verification coverage, it has also shifted the balance between the numerous discovered vulnerabilities and the speed at which they can be verified and fixed, creating new challenges.


Such changes in balance may ripple throughout security, and by leveraging and complementing each other's strengths, AI and humans are expected to establish highly effective security operations.


Who owns powerful AI ? Democratizing AI Design as Demonstrated by OpenMythos


1.1. Overview

A 22-year-old individual developer attempted to rebuild Anthropic's private AI model "Claude Mythos Preview" (hereafter, Claude Mythos) based on a third-party academic paper, launching it as the open-source project "OpenMythos" on GitHub (a platform for developers to manage and share program source code). Published above.[1]


Within the platform community, it has attracted significant interest and support from developers and researchers.[2]


Figure 1: OpenMythos explanatory image (from developer Kye Gomez's X post)[3]
Figure 1: OpenMythos explanatory image (from developer Kye Gomez's X post)[3]

1.2. What is OpenMythos?


About OpenMythos

Although the names are similar, OpenMythos and Claude Mythos are not officially related. Before the announcement of Claude Mythos, due to a misconfiguration in Anthropic's CMS (Content Management System), the company's unpublished data (including those related to Claude Mythos) became accessible to anyone.[4]


OpenMythos development did not take advantage of this leak. Additionally, Anthropic has not published technical papers or materials regarding the model structure of Claude Mythos, and Kye Gomez, developer of OpenMythos, inferred the design of Claude Mythos by combining the content of research introduced in several existing papers (not involving Anthropic).[5]


The source code is published on GitHub and can be obtained, used, and modified by anyone free of charge. Incidentally, the paper on design philosophy adopted by Gomez states that it can deliver performance equivalent to larger AI models with about half the number of parameters (indicators of AI processing scale) compared to conventional ones.[6]


This suggests that advanced AI could be powered even on cheaper, lower-spec hardware in the future.


Currently, no benchmark evaluation for OpenMythos (a method that compares and quantifies model performance using datasets and metrics) has not been confirmed. Additionally, the project does not provide the trained data called "weights" necessary for AI to actually operate,[7]


Instead, it publishes recommended datasets, which users can use to train OpenMythos.[5]


Developer Profile

The developer, Mr. Gomez, is a 22-year-old from Miami, Florida, USA. He began programming at age 10, and at 13, he created his first AI model, aiming to hack his mother's Gmail account and obtain code codes that could be used in the game's online store.[9]


Currently, he is a researcher focused mainly on AI and serves as CEO of Swarms, an AI company founded several years ago.[10]



Evaluation of OpenMythos

In the past, Gomez's published works were criticized within the machine learning community for altering and releasing others' research without citing sources.[5]


In the release of OpenMythos, Gomez explicitly cited the papers he referenced and is not claiming unfounded originality. However, there is no guarantee that the content of this project will accurately reflect the structure of Claude Mythos; as he himself says, it is only "publicly available research and reasoning."[2]


It is important to note that it is based on these principles.[5]


The American business magazine Forbes reported that the significance lies not in whether OpenMythos matches Anthropic's design but in the fact that the internal design of AI models can be reconstructed from public information, and praised the project's release for providing researchers with concrete verification materials.[7]


In fact, this kind of evaluation is reflected in the response on GitHub, where OpenMythos has gathered over 10,000 stars (indicators of interest and support from GitHub users), indicating significant attention from researchers and developers.[2]


Figure 2: Number of GitHub stars (red frame, as of May 21, 2026)[2]
Figure 2: Number of GitHub stars (red frame, as of May 21, 2026)[2]

Anthropic's Perspective

At this time, Anthropic has not made any comments regarding OpenMythos. On the other hand, the company has expressed a view that the time it takes for open source and Chinese developers to reach capabilities equivalent to Claude Mythos is generally about 6 to 12 months.[11]


1.3. Democratization of AI Design Philosophy

What should be noted when considering changes surrounding the development of cutting-edge AI is the current ability to reconstruct "who can reconstruct their design philosophy by how fast and to what extent." OpenMythos presents cutting-edge AI design methods that have traditionally been treated as private in limited environments such as large IT companies, but the open source community enables rapid experimentation and validation. This has led to a movement to share knowledge related to AI model design in a practical and trial-based manner, which is significant in that the scope of research and development is expanding.[7]


1.4. Summary

The emergence of OpenMythos has shown that the fundamental question of "who can have powerful AI" may be changing. It is becoming a natural trend that the design of cutting-edge AI, which was previously mainly limited to large IT companies, will expand the possibilities for free discussion and verification within open source communities unaffected by government regulations or corporate intentions. In the future, as the use of powerful AI expands, it will be necessary to steer the company and organizations to update their defense postures.

[1] Source: MarkTechPost 'Meet OpenMythos: An Open-Source PyTorch Reconstruction of Claude Mythos Where 770M Parameters Match a 1.3B Transformer'


[2] Source: GitHub 'kyegomez/OpenMythos'https://github.com/kyegomez/OpenMythos



[4] Source: Awesome Agents 'Anthropic's Mythos Model Exposed by CMS Misconfiguration'https://awesomeagents.ai/news/anthropic-mythos-capybara-leak/


[5] Source: Awesome Agents 'OpenMythos Recasts Claude Mythos as Looped MoE Transformer'https://awesomeagents.ai/news/openmythos-recurrent-depth-transformer/


[6] Source: Sandy Research 'Parcae: Doing More with Fewer Parameters using Stable Looped Models'https://sandyresearch.github.io/parcae/


[7] Source: Forbes 'Did A 22-Year-Old Dropout Reverse-Engineer The World's Scariest AI?'


[8] Source: GitHub 'kyegomez/OpenMythos - Recommended Training Datasets'https://github.com/kyegomez/OpenMythos/blob/main/docs/datasets.md


[9] Source: Refresh Miami '18-year-old Miamian Kye Gomez is developing AI to make life less boring'


[10] Source: Kye Gomez 『Kye Gomez | AI Researcher and Founder of Swarms.ai』


[11] Source: SecurityWeek 『Chinese Cybersecurity Firm's AI Hacking Claims Draw Comparisons to Claude Mythos』


Proximity Attack on Taiwan High-Speed Rail


2.1. Overview

In Taiwan, a 23-year-old man used commercially available software-defined radio (SDR) equipment to hack the wireless communication system of the Taiwan High Speed Rail (THSRC). The attackers sent a fake "emergency alert" signal, stopping four trains and causing a total delay of 48 minutes.[12]


This incident demonstrated that close-range attacks using commercial equipment can physically halt critical infrastructure, exposing serious vulnerabilities in railway operational technology (OT).[13]


2.2. About the Attackers

The attacker is a 23-year-old Taiwanese university student reported to be an enthusiast of wireless communications.[14]


Regarding the attack, he claims that he "accidentally misoperated the radio,"[15]


During the police investigation, 11 professional software radios, PCs, and other items were seized from his home. Given these circumstances, it is highly likely that signal decryption was intentionally performed.


Figure 3: Seized equipment (all easily available commercial items)[16]
Figure 3: Seized equipment (all easily available commercial items)[16]


Software-defined radios are generally commercially available and feature flexible adaptation to different wireless communication methods and frequencies by connecting to a PC and rewriting the software. While highly flexible, it enables interception of surrounding wireless communications, signal analysis, and even the generation and transmission of signals that mimic specific communication methods.


In this case as well, the attackers exploited these characteristics to analyze and reproduce Taiwan High Speed Rail's wireless communication protocols and parameters, intentionally sending signals disguised as legitimate communication.


Furthermore, investigations confirmed that the attacker was capable of receiving and analyzing communication frequencies not only for Taiwan High-Speed Rail but also for fire departments and airport MRT (subway stations), revealing that they had broad access to highly public wireless infrastructure.[12]


2.3. Operational Deficiencies in Wireless Infrastructure

The main cause of this incident lies in long-standing operational management deficiencies in wireless communications. In particular, in wireless communication areas accessible directly from outside, the lack of proper management of authentication information and settings allowed close-range attacks to infiltrate and directly led to the success of the attack.


Operational Shortcomings

By transmitting radio waves within the wireless communication area, it becomes possible to send signals to communication systems, but this alone is difficult to compromise the system. However, it has been pointed out that the system used by Taiwan High Speed Rail may have not changed the parameters and authentication information used for communication for about 19 years.[17]


Under such circumstances, information obtained through observation and analysis of communications accumulated, making it easier to grasp communication methods and related settings based on this.

Furthermore, the attacker obtained communication parameters used for authentication from a friend (accomplice),[18]


It is believed that by combining this information, the conditions required for authentication could be reproduced, allowing the communication to be disguised as legitimate communication. These issues were mainly caused by operational management issues, and ideally, they could have been prevented by regular updates and proper management of credentials and communication parameters.


Impact of the Incident


On May 4, during deliberations in the Transportation Committee of Japan's Diet, the Legislative Yuan, a lawmaker stated, "The wireless communication system for high-speed rail is already at the highest level among Taiwan's transportation systems. If even high-speed rail is hacked, wouldn't Taiwan's conventional railway become even more vulnerable?" He called for the need to review the timing of radio system upgrades and procedures for equipment storage. In response, Taiwan's Ministry of Transportation explained that inspections are being conducted targeting Taiwan High Speed Rail and Taiwan Railways, and that they plan to request responses from subway operators as well.[19]



2.4. Summary

Depending on the management conditions, wireless communication has a structure that can be vulnerable to attacks from close range. This case demonstrates that, under such characteristics, if authentication information and communication settings are not properly managed, security functions may not be fully utilized, which can lead to receiving signals transmitted from outside. Similar wireless communication systems are widely used in the railway sector, including Japan, but these characteristics are not limited to railways but are also seen in other fields utilizing wireless communication. This case suggests cross-disciplinary risks.

[12] Source: Taipei Times 'Student's hack prompts THSRC review'https://www.taipeitimes.com/News/taiwan/archives/2026/05/05/2003856781


[13] Source: Dark Reading 'Taiwan Bullet Train Hack Highlights Cybersecurity Gaps in Rail Systems'


[14] Source: CNY News 'High-Speed Rail Radio Hacked with Accomplice!' Male college student junior provides key parameters, emergency button triggered outside Wuri Station exposed'https://www.knews.com.tw/news/85322068F027005083050EC2DDFC3DE4


[15] Source: SETN Sanlih News Network 'Eavesdropping on High-Speed Rail Radio! 23-Year-Old Male College Student 'Accidentally Presses Emergency Button,' Causing Three Trains to Stop Abruptly—Shocking Background Revealed'https://www.setn.com/News.aspx?NewsID=1831451


[16] Source: FTV News Network 'Called a 'National-Level Hacker'! Male College Student Hacks High-Speed Rail and Taipei Metro Forcing Train Stoppages https://www.ftvnews.com.tw/news/detail/2026430C13M1


[17] Source: ETtoday News Cloud 'High-Speed Rail Radio Verification Channel 7 Still Stolen; Legislator: System Hasn't Been Replaced for 19 Years'https://www.ettoday.net/news/20260430/3158010.htm


[18] Source: Central News Agency 'High-Speed Rail Radio Theft Case: Prosecutors Arrest 20-Year-Old Accomplice and Released on Bail


[19] Source: Liberty Times 'High-Speed Rail Radio Interrupted and Delayed; Ministry of Transportation: Review Within One Month'https://news.ltn.com.tw/news/Taipei/breakingnews/5424768

Structural Transformation of Bug Bounties Driven by AI


3.1. Overview

Google in the US is reviewing the Bug Bounty system. This is largely due to the significant changes in vulnerability detection methods brought about by the spread of AI. With the use of generative AI, investigations have become more automated and advanced, greatly expanding the scope of verification that previously relied on manual labor, and significantly shortening the time to discovery. In light of this situation, companies like Google, which have introduced bug bounty, are shifting their system designs to adapt to the new era, based on comprehensive AI-based investigations.


Figure 4: Announcement regarding Google's bug bounty review[20]
Figure 4: Announcement regarding Google's bug bounty review[20]


3.2. What is a bug bounty?

A bug bounty is a system in which external security researchers or white-collar hackers discover and report vulnerabilities in an organization's products or services, and a reward is paid in return. The purpose of this system is to identify vulnerabilities before they can be exploited by attackers and quickly fix them to reduce security risks. Because diverse global expertise can be utilized, vulnerabilities that are often overlooked by conventional diagnostic methods can be expected, and the performance-based model offers the advantage of high cost-effectiveness.[21]


This system has been established on the premise that researchers with advanced expertise take the lead in discovering and verifying vulnerabilities.[22]



3.3. The Impact of AI

AI has dramatically improved the efficiency of vulnerability investigations and greatly expanded the analytical process that previously relied on a handful of skilled security researchers. Nowadays, advanced automation technologies using AI are being widely applied to detect vulnerabilities. For example, it is said to be possible to scan vast amounts of code in a short time and detect anomalies and misconfigurations that were previously overlooked, which greatly contributes to reducing dependence on personnel and standardizing quality.[23]


With AI assistance, more researchers can efficiently perform everything from vulnerability identification to verification and reporting, expanding the base of bug bounty participants.


However, this change also brings new structural challenges. This is because the acceleration of AI-driven "discovery and reporting" is overwhelming human processing capabilities for "verification and correction." Massive resources are being drained by verifying the flood of reports, leading to 'triage fatigue'—where just identifying the most critical bugs that truly need to be addressed—and 'fixing process bottlenecks'—where developers' ability to fix them cannot keep up with the number of cases deemed valid, leaving vulnerabilities unresolved.[24], [25]


The evolution of AI is driving a shift from discovering large numbers of vulnerabilities to the next stage: how to fix and manage them quickly and efficiently. This shift in the balance of discovery and remediation in vulnerability investigations has led to a redesign of bug bounties that accept these outcomes.



3.4. Responses by Each Company

To adapt to major changes surrounding bug bounty, organizations implementing this system are required to redesign their evaluation criteria and optimize their operations.


Google

Android: Google sets rewards in stages according to the severity of the vulnerability. With this review, especially for vulnerabilities in Google Pixel's security chip that involve "persistence," where attacker access and control continue after a breach, the bounty cap has been raised from $1 million to $1.5 million.[27]


Behind these substantial rewards lies the premise of extensive AI investigations, while focusing on vulnerabilities that are difficult to discover with automated AI and require human contextual understanding and analysis.


Chrome: We shifted our policy from detailed explanations to prioritizing concise reports with reproducible evidence. Additionally, as AI made it easier to demonstrate certain attack techniques, the bonus system targeting those attacks was abolished.


Figure 5: List of maximum rewards for Android vulnerability reward programs
Figure 5: List of maximum rewards for Android vulnerability reward programs[28]

HackerOne


HackerOne, known as a leading platform for bug bounty, operated the Internet Bug Bounty (IBB) vulnerability bounty program targeting major open source software (OSS), but has stopped accepting new applications. The background is that while the speed of vulnerability discovery has greatly improved through the use of AI, OSS fixes have lagged behind, disrupting the balance between the two. Originally, the IBB aimed to ensure reliable fixes by discovering vulnerabilities and achieving continuous security improvements, but the system is currently undergoing revision.[29]


GitHub


GitHub, a platform for developers to manage and share program source code, announced a policy to reward reports that do not cause significant security impacts but that lead to code or documentation fixes will be rewarded with novelty merchandise rather than bounty rewards. The aim is to concentrate bounty resources on high-severity vulnerabilities.[30]



3.5. Summary

Until now, bug bounties have tended to emphasize the number of vulnerability reports, but with the spread of AI, this assumption is changing significantly. Certainly, AI has dramatically increased vulnerability discoveries, but at the same time, fixes and management have lagged behind, creating new bottlenecks in determining the importance and priority of response. Such changes may ripple beyond the bug bounty field into the entire security sector, potentially causing various disruptions.


Additionally, for the time being, AI will be responsible for detection and analysis in security operations, while humans will evaluate and prioritize, expanding the division of roles in various aspects. On the other hand, as AI evolves, there may be more cases where AI surpasses human capabilities in evaluation and decision-making. In such circumstances, building flexible role divisions that leverage the strengths of both sides and establishing highly effective security operations will become a key challenge going forward.

[20] Source: Google 'Evolving the Android & Chrome VRPs for the AI Era'https://bughunters.google.com/blog/evolving-the-android-chrome-vrps-for-the-ai-era


[21] Source: Mitsui & Co. Secure Direction Co., Ltd. 'Bug Bounty'https://www.mbsd.jp/glossary/bug_bounty


[22] Source: Technology Org 'Bug Bounty Schemes Buckle Under Flood of AI-Generated Junk Reports'


[23] Source: Secure Innovation '[2025 Latest] How Far Has AI-Based Vulnerability Diagnosis Evolved?'


[24] Source: Dark Reading 'AI-Led Remediation Crisis Prompts HackerOne to Pause Bug Bounties'


[25] Source: HackerOne 『The Vulnerability Apocalypse Is a Remediation Crisis』


[26] Source: SecurityWeek 『Google Adjusts Bug Bounties: Chrome Payouts Drop as Android Rewards Rise Amid AI Surge』


[27] Source: Google 'Evolving the Android & Chrome VRPs for the AI Era'https://bughunters.google.com/blog/evolving-the-android-chrome-vrps-for-the-ai-era


[28] Source: Google 'Android and Google Devices Security Reward Program Rules'


[29] Source: HackerOne 'Inernet Bug Bounty - Program guidelines' (Submission page)


[30] Source: GitHub 'Raising the bar: Quality, shared responsibility, and the future of GitHub's bug bounty program'

  1. Disclaimer

Please note that while we do our best to ensure that the content of this article is accurate, we do not guarantee the content and do not compensate for any damages or losses incurred as a result of the use of this article. If you have any inquiries such as typographical errors, content errors, or other points in the article, please contact us at the following address.


Inquiries

NTT Security Japan Corporation

Professional Services OSINT Monitoring Team

 
 
bottom of page