Blogs

As generative AI and AI Agents rapidly enter enterprise environments, AI Governance is becoming an increasingly important part of enterprise risk management. Many organizations have begun establishing AI governance policies that define which AI tools employees may use, what types of data must not be entered into AI systems, and what cybersecurity and compliance requirements AI applications must follow. However, establishing policies is only one part of AI Governance. Policies define how an organization expects AI to be used, but effective governance also requires visibility into whether actual AI usage aligns with those requirements. For enterprises, this raises a practical question: What AI is actually being used on endpoints? If an organization knows what rules it has established but lacks visibility into how AI is actually deployed and used, an Endpoint Visibility Gap may emerge within its AI Governance framework. AI Governance Is More Than Establishing Policies Comprehensive AI Governance encompasses policies, roles and responsibilities, risk management, data governance, security controls, continuous monitoring, and auditing. AI governance policies can establish the fundamental principles for how AI should be used within an organization. However, to determine whether these requirements are actually being followed, enterprises also need visibility into their real-world environments. For example: Which generative AI tools or AI Agents are employees actually using? Are there any unauthorized AI tools in use? Do AI Agents have access to sensitive files or critical systems? Are there insecure AI-related configurations on endpoints? Without visibility into the actual environment, organizations may face a gap where policies are in place, but actual AI usage remains unseen. Why Are Endpoints Critical to AI Governance? Traditionally, enterprise software adoption has typically gone through IT procurement, deployment, and access management processes. Generative AI, however, has changed this model. Employees can directly access cloud-based AI services, browser tools, and IDE extensions, or even install AI Agents capable of autonomous actions. As a result, endpoints are becoming an important environment where enterprise AI usage—and its associated risks—actually materialize. For example, employees may install unauthorized AI tools on corporate devices or grant AI Agents access to local files, browsers, and other applications. These activities may not be visible through governance policies alone, yet they can directly affect an organization’s data access and risk exposure. Even when an organization has explicitly prohibited certain activities, it is difficult to verify whether those policies are being followed without visibility into the actual deployment, configurations, and permissions on endpoints. This gap—where governance policies define the rules but cannot reveal actual AI usage—is the Endpoint Visibility Gap that organizations need to address when implementing AI Governance. Moving AI Governance from “Policy” to “Verifiable Practice” To narrow this gap, organizations can inventory AI applications and Agents on endpoints, review their configurations and permissions, and identify risks based on governance policies. Endpoint assessment does not cover every aspect of AI Governance. Organizations still need policies, data governance, compliance, model risk management, and clear accountability. However, endpoint visibility provides an important foundation by showing how AI is actually used in enterprise environments. TeamT5 ThreatSonar Plus supports this effort through AI Agent Detection, configuration assessment, and vulnerability assessment , helping security and IT teams identify endpoint risks and prioritize remediation. AI Governance should address not only “What rules have we established?” but also “Does our actual environment comply with those rules?” Connecting governance requirements with actual endpoint deployments, configurations, and permissions helps narrow the gap between AI governance policies and real-world AI usage .

Generative AI and AI Agents are rapidly becoming part of enterprise environments. Employees may use AI services such as ChatGPT, Claude, and Gemini, while development or IT teams may deploy various AI Agents across the organization. As enterprise AI adoption becomes increasingly diverse, the first question organizations need to address is no longer whether they should use AI, but rather: Do we actually know which AI Agents are being used across our organization? Establishing this visibility is a fundamental step in putting AI Governance into practice. What Is Essential to AI Governance? AI Governance refers to the policies, responsibilities, and management mechanisms organizations establish to govern AI adoption, usage, and associated risks. It encompasses areas such as data governance, cybersecurity, privacy, regulatory compliance, model risk, and continuous monitoring. AI Governance therefore cannot be achieved simply by deploying a single tool. It requires a cross-functional management framework. Regardless of the governance approach an organization adopts, however, it must first answer a fundamental question: What AI is actually being used within the organization? Without a clear understanding of AI assets and usage, organizations may face blind spots when conducting risk assessments, establishing policies, and implementing security controls. Why Should Enterprises Inventory AI Agent Usage First? AI adoption often moves faster than internal governance processes can keep up. In addition to officially approved AI services, employees may independently sign up for generative AI platforms, install AI applications, or authorize AI Agents to access files, browsers, and other corporate resources. This can create a Shadow AI problem similar to Shadow IT. Unmanaged AI applications may introduce risks such as sensitive data leakage, inappropriate access permissions, malicious instructions, and exposure to third-party services. Building an AI Inventory is therefore an important foundation for effective AI Governance. What Should Enterprises Include in an AI Agent Inventory? The first step is to establish a basic AI Inventory. Organizations can document information such as the AI tool or service name, department using it, purpose of use, types of data involved, deployment method, and responsible personnel. This helps answer a basic question: Who is using which AI, and for what purpose? The second step is to gain visibility into the AI Agents actually running on endpoints. Policy reviews and employee surveys can only reveal AI usage that the organization already knows about. If employees independently install AI applications, browser extensions, or AI Agents, IT and security teams may still lack complete visibility. The third step is to classify and prioritize risks. For example: Does the AI have access to confidential documents? Can it read or write files, execute programs, or connect to external services? Do its permissions exceed actual business requirements? Organizations can prioritize subsequent governance and security measures based on data sensitivity, access privileges, and usage scenarios. Build the Foundation for AI Governance by Making AI Visible AI Governance covers policies, processes, compliance, and cybersecurity. While an AI inventory is only one part of governance, visibility into actual AI usage is essential for assessing and managing risks. TeamT5 ThreatSonar Plus helps organizations identify AI Agents and related risks from an endpoint security assessment perspective, providing greater visibility to support risk assessment and governance decisions. As AI adoption grows, organizations can start by identifying which AI tools and Agents are being used, where they operate, and which corporate resources they can access . This visibility provides a practical foundation for effective AI risk management and governance.

Campaign Snapshot Campaign Timeframe: Jan 2026 to May 2026 Delivery Methods: Phishing Victim Country: Taiwan Phishing Kit: Darcula Actor Assessment: Chinese-speaking Executive Summary In May, we intercepted a phishing campaign harvesting Taiwanese credit card data. The threat actor leveraged a fake Taiwanese e-invoice platform to lure victims into submitting their credit card information through phishing pages. Our technical analysis found that the campaign encrypts victim data using the Rabbit encryption algorithm. The Rabbit encryption algorithm is the same encryption mechanism previously observed in the Darcula phishing kit. Based on this technical overlap, we assess with high confidence that the campaign was conducted using the Darcula phishing kit. We list all the malicious URLs in the IoCs section below. 1. Phishing Emails In this campaign, the actor delivered the phishing emails from a compromised email account belonging to Korea University[1]. In these emails, the actor impersonated the Taiwanese e-commerce platform MOMO and instructed victims to click the malicious links to verify their e-invoice. The actor prepared several malicious URLs that direct victims to the fake e-invoice platform: - https://0023.ehrscripts.com/ - https://av11.pdjekqa.online/ - https://einvoiceg.com/gov/ - https://einvoicegs.com/gov/ 2. Fake Taiwanese E-invoice Platform At the time of our analysis, we were only able to access two URLs https://einvoiceg.com/gov/ and https://einvoicegs.com/gov/ . The two URLs show the page identical to the legitimate e-invoice platform. However, the icon of the page displays the logo of Taiwanese commercial bank CTBC rather than the logo of Ministry of Finance used by the legitimate site. Figure 1: The fake e-invoice platform While the fake e-invoice platform requires a phone number to sign in, we found that any number entered resulted in a successful login. Once logged in, the page requests credit card information so that the e-invoice can be linked with the credit card. This lure is effective because e-invoices in Taiwan are eligible for a government-run lottery, and residents routinely retain their invoices in the hope of winning a prize. Figure 2: The fake platform that requires phone number to sign in Our research shows that the phishing page performs basic validation of the credit card information. Notably, the error messages shown for invalid input are displayed in Simplified Chinese. Figure 3: Simplified Chinese error messages After the credit card information is submitted, the page redirects to a second page requesting two-factor authentication (2FA) information. We assess that this data is likely synchronized to the phishing kit's backend server in near real time. Figure 4: The page requesting 2FA authentication 3. Relations to Darcula Our technical analysis of the phishing kit’s source code reveals that the victim data is encrypted with the Rabbit algorithm. The algorithm has been previously documented in the Darcula analysis.[2] Specifically, we identify a file app/chunk/DgZYu39z.js that contains an encryption and decryption mechanism sharing the same structure as the Darcula phishing kit. Therefore, we assess with high confidence that the campaign was deployed using the Darcula phishing kit. Notably, the actor may have used AI during development, as we identified numerous Simplified Chinese strings along with emoji characters in the source code. Footnotes [1] Korea University https://www.korea.ac.kr/sites/ko/index.do [2] Exposing Darcula: a rare look behind the scenes of a global Phishing-as-a-Service operation https://www.mnemonic.io/resources/blog/exposing-darcula-a-rare-look-behind-the-scenes-of-a-global-phishing-as-a-service-operation

SEMI E187 establishes a cybersecurity baseline for semiconductor manufacturing equipment and defines the security capabilities equipment should have before entering a wafer fabrication facility. However, for many equipment suppliers and semiconductor manufacturers, the greatest challenge is not understanding the standard but translating its requirements into routine security assessment procedures. When assessments still rely on manually reviewing documents and checking configurations item by item, the process becomes time-consuming and difficult to perform consistently. Establishing a repeatable and measurable equipment security assessment process is therefore essential to implementing SEMI E187 effectively. Step 1: Build a Comprehensive Equipment Asset Inventory Effective security management begins with a clear understanding of the equipment and its assets. Start by identifying the equipment’s operating system versions, installed software, firmware versions, and network services. Confirm whether each component is still supported by its original vendor and establish a complete asset inventory as the foundation for subsequent risk assessments. Step 2: Assess Compliance with SEMI E187 Requirements After completing the asset inventory, assess the equipment against the core requirements of SEMI E187, including: Whether the operating system is still supported and regularly updated Whether network communications use encryption Whether unnecessary ports and services have been disabled Whether vulnerability remediation and malware protection capabilities are in place Whether account, privilege, and access controls have been properly implemented Whether complete logs are retained for auditing purposes Together, these controls constitute the fundamental security capabilities equipment should have before deployment and serve as important evidence during SEMI E187 validation. Step 3: Replace Manual Judgment with Automated Assessments In practice, many risks cannot be verified through documentation alone. For example: Has the operating system reached the end of support? Are default accounts still in use? Is HTTP traffic transmitted without encryption? Are high-risk ports such as VNC port 5900 exposed? Do communications lack encryption or authentication? Relying entirely on manual verification increases the likelihood of omissions and makes it difficult to maintain consistent assessment standards. Automated assessment tools can directly inventory equipment configurations, correlate findings with vulnerability intelligence, and convert previously ambiguous risks into measurable assessment results. This significantly improves both efficiency and accuracy. Step 4: Establish Risk Classification and Reporting Equipment assessments should not produce only a “pass” or “fail” result. Organizations should classify identified weaknesses according to severity and summarize them using risk levels such as Critical, High, Medium, and Low. Reports should also document remediation status, Windows hotfixes, software and firmware versions, and compliance gaps. This creates a structured record that can support audits, remediation planning, and management decision-making. Step 5: Move from One-Time Assessments to Continuous Management The purpose of SEMI E187 is not merely to complete a one-time validation. It is to establish ongoing cybersecurity governance for semiconductor manufacturing equipment. New risks may emerge throughout equipment delivery, deployment, production, and maintenance due to software updates, configuration changes, or newly disclosed vulnerabilities. Organizations should therefore establish periodic assessments, continuous monitoring, and incident response mechanisms. Equipment security should become part of routine operational management rather than a temporary activity conducted only before validation. Strengthen Cyber Resilience from Assessment to Protection SEMI E187 provides a common language for equipment cybersecurity, but its real value comes from converting the standard into an actionable assessment process. By integrating asset inventory, configuration assessment, vulnerability analysis, and continuous monitoring, organizations can improve validation efficiency while establishing an equipment security management framework that is measurable, traceable, and continuously improved. When security assessments become part of daily operations, SEMI E187 is no longer simply a compliance requirement. It becomes an important foundation for strengthening the resilience of the semiconductor supply chain. Is Your Equipment Compliant with SEMI E187? Whether you are preparing for SEMI E187 validation or seeking a faster way to understand the cybersecurity posture of your equipment, TeamT5 can help you establish an equipment security assessment process aligned with SEMI E187. Through automated asset inventory, vulnerability analysis, configuration assessment, and compliance reporting, TeamT5 helps reduce the cost of manual inspections while improving the efficiency and consistency of pre-deployment equipment assessments. Discover ThreatSonar Plus and learn how to transform SEMI E187 requirements into a sustainable equipment security management process. Notes This article references the SEMI E187 standard for educational and explanatory purposes only. The copyright of the standard belongs to SEMI, Semiconductor Equipment and Materials International. Official SEMI E187 requirements and interpretations should be based on the latest version published by SEMI.

As cybersecurity requirements across the semiconductor supply chain continue to rise, SEMI E187 is becoming an increasingly important standard for both equipment suppliers and semiconductor fabs. For equipment manufacturers, E187 is not simply about passing a validation process. It demonstrates that equipment has essential cybersecurity capabilities in place before delivery, helping reduce deployment risks across the supply chain. How should equipment suppliers prepare for SEMI E187 validation? What Does SEMI E187 Validation Focus On? SEMI E187 focuses on the cybersecurity capabilities of fab equipment. It applies primarily to Windows- or Linux-based computing devices embedded in the equipment. The standard requires suppliers to provide relevant cybersecurity information and enables fabs to verify whether the equipment meets established security baselines. Key assessment areas include operating system security, network security, endpoint protection, access control, and logging. SEMI E187 Validation Checklist 1. Operating System Management Confirm that the operating system used by the equipment is still supported by the original vendor and has not reached end of life (EOL). Establish comprehensive patch and update management procedures to prevent unsupported or unmaintained operating systems from remaining in use. 2. Network Security Configuration Verify that the equipment uses encrypted communications and inventory all enabled network services and ports. High-risk services such as Telnet and FTP should be disabled. Only essential communication protocols should remain enabled to minimize the equipment’s attack surface. 3. Endpoint Protection Establish a vulnerability remediation process, confirm that the equipment has undergone malware scanning, and ensure that appropriate anti-malware protection is available. System configurations should also be hardened by restricting USB usage, disabling unnecessary services, and limiting local software installation privileges. 4. Account and Privilege Management Disable default accounts, establish a password policy, and avoid the use of shared accounts. Access privileges should be assigned according to user roles to ensure that all access to the equipment can be attributed and audited. 5. Logging and Audit Capabilities The equipment should retain comprehensive logs covering login activity, configuration changes, system errors, and other relevant events. These logs provide critical evidence for future audits, incident investigations, and compliance verification. Documentation Alone Is Not Enough—Equipment Security Must Be Verifiable Many organizations assume that providing the necessary documentation is sufficient to complete the validation process. In practice, however, SEMI E187 places greater emphasis on whether the equipment has cybersecurity capabilities that can be independently verified. For example: Does the equipment contain known vulnerabilities? Are default accounts still enabled? Are high-risk ports exposed? Does the equipment use unencrypted communications? Relying entirely on manual verification can be time-consuming and may leave critical risks undetected. As a result, a growing number of equipment suppliers are introducing automated assessment tools. Through asset inventory, vulnerability correlation, configuration assessments, and compliance reporting, organizations can transform processes that previously depended on manual judgment into measurable and traceable assessment workflows. This improves both the efficiency and consistency of SEMI E187 validation. Conclusion The purpose of SEMI E187 is not to create additional burdens for businesses. It is intended to establish a common cybersecurity baseline for semiconductor manufacturing equipment. For equipment suppliers, implementing a standardized assessment process at an early stage can: Improve validation efficiency Reduce supply chain risks Strengthen customer confidence in equipment security Establish a foundation for ongoing equipment cybersecurity governance Need to meet SEMI E187 compliance requirements within a limited timeframe? Contact TeamT5 to learn how ThreatSonar Plus can help your organization establish an automated cybersecurity assessment process for semiconductor equipment. Disclaimer This article references the SEMI E187 standard and is intended solely for educational and explanatory purposes. Copyright for the standard belongs to SEMI—Semiconductor Equipment and Materials International. Official SEMI publications should be regarded as the authoritative source for SEMI E187 requirements and interpretations.

As Generative AI technology continues to evolve at an unprecedented pace, enterprise AI adoption is undergoing a critical paradigm shift. We are moving beyond the era of chatbots that simply respond to user prompts and entering the age of AI Agents—systems capable of independently planning tasks, invoking tools, and directly executing system commands. While this technological revolution offers tremendous productivity gains, it also introduces entirely new cybercybersecurity blind spots. Without proper governance, once AI Agents begin acting as “autonomous operators” on enterprise endpoints, traditional endpoint cybersecurity mechanisms face unprecedented challenges. 1. Three Common Enterprise AI Agents and Their Endpoint Cybersecurity Risks According to statistics from the TeamT5 support service team, the following three AI Agents are among the most commonly observed on enterprise endpoints. Understanding their characteristics can help organizations identify potential cybersecurity vulnerabilities. 1. OpenAI Codex: A New Blind Spot in Software Supply Chain cybersecurity Positioning : Codex is deeply integrated into developers’ IDEs (Integrated Development Environments). It can autonomously analyze project context and automatically complete or modify code. Risk : In addition to the possibility of source code being passively uploaded during project analysis, credential-related vulnerabilities disclosed in early 2026 demonstrated that if Codex is compromised through privilege escalation, attackers may be able to move laterally into an organization’s software hosting platforms. This could allow them to implant malicious backdoors directly into source code repositories, threatening the cybersecurity of the entire software supply chain. 2. Claude: A Major Source of “Shadow AI” in Reasoning and Analytical Workflows Positioning : With its strong logical reasoning capabilities, long-context processing, and safety alignment, Claude is frequently used as a core tool in enterprise automation workflows. Risk : It is also one of the AI tools most commonly used by employees outside formal governance processes to handle confidential documents, making it a major source of “Shadow AI.” When granted access to internal corporate APIs or email systems, Claude may become vulnerable to Prompt Injection attacks, potentially causing sensitive internal information to be unintentionally transmitted to external users. 3. Cline (Claude Dev): The Endpoint “Autonomous Operator” and a Blind Spot in Behavioral Monitoring Positioning : Cline is a highly popular autonomous AI coding agent among developers. Integrated directly into IDEs such as VS Code, it can independently plan task steps, read and write local files, execute terminal commands on endpoints, and even launch browsers to perform application testing. Risk : Cline is granted a high degree of autonomous control. Its system activities—such as reading or writing files and executing commands—appear indistinguishable from normal VS Code development behavior under traditional process-level monitoring. If Cline is manipulated through malicious prompts, it may execute unintended code, potentially resulting in remote code execution (RCE) or the deletion of critical files. 2. How to Identify AI Agents in Your Environment As enterprises face emerging endpoint cybersecurity threats in the AI era, they need appropriate tools to understand how AI Agents are being used across their environments. 1. ThreatSonar’s Approach to AI Discovery Through ThreatSonar’s Threat Hunting interface, cybercybersecurity teams can identify relevant endpoint activity without disrupting endpoint operations. EDR-based real-time detection and scheduled scanning collect essential information about processes and files on endpoints. The primary investigation methods include: File and Attribute Search : Identify characteristics associated with commonly used AI tools. Event Log and Command Search : Search for known AI Agent keywords and detect relevant commands in real time. Connection IP Analysis : Monitor whether endpoint processes are transmitting data to known AI service API endpoints, such as OpenAI or Anthropic. Examples include: Using Threat Hunting to identify execution characteristics associated with OpenAI. Using Threat Hunting to identify execution characteristics associated with Claude. Using Threat Hunting to identify execution characteristics associated with Cline. 2. Limitations of Investigation Mechanisms for Known AI Agents However, the investigation methods described above are primarily effective against known AI Agents. When dealing with unauthorized “Shadow AI” deployments or highly autonomous AI Agents, normal activity can be difficult to identify unless explicitly malicious commands are executed. Difficulty identifying malicious intent within legitimate behavior : For unknown AI Agents, unless they execute highly obvious malware or known malicious commands, activities such as reading and writing files, executing system commands, and calling APIs appear entirely legitimate under traditional process-level monitoring. As a result, conventional mechanisms may struggle to identify suspicious behavior at an early stage. Lack of AI behavioral context : Traditional cybersecurity tools cannot understand the relationship between natural-language prompts and the system commands generated from them. They therefore cannot determine whether a particular command reflects the user’s actual intent or whether it is the result of an AI Agent being manipulated through Prompt Injection. 3. Comprehensive cybersecurity Governance: ThreatSonar Plus Visibility and Endpoint Defense To address the new threats introduced by AI Agents, TeamT5 has launched ThreatSonar Plus, a comprehensive endpoint cybersecurity assessment platform designed to counter the risks of Shadow AI and uncontrolled AI Agents through the following capabilities. 1. Core Advantage: Extending Visibility from System Processes to AI Behavior ThreatSonar Plus introduces the following key capabilities: Behavioral Visibility : Systematically assess the presence and activity of AI Agents on endpoints, allowing administrators to clearly understand Agent configurations and eliminate Shadow AI blind spots across the environment. Command-level Detection : ThreatSonar Plus focuses on understanding the actual commands executed by the AI agent. It analyzes the command patterns to identify potentially abnormal or unexpected behaviors, thus grasping the operational outline of the AI agent. 2. Core Advantage: Targeted Defense Against OWASP Top 10 Risks ThreatSonar Plus provides targeted defensive mechanisms against key OWASP-related threats: Prevent Goal Hijacking and Tool Misuse : Identify whether an AI Agent is invoking unusual “skills” or tools, helping prevent malicious manipulation of Agent behavior. Detect Identity Anomalies and Sensitive Data Exposure : Accurately identify where keys, credentials, and sensitive information are stored on endpoints, reducing the risk of unauthorized access by AI Agents. Strengthen Supply Chain and Code Execution cybersecurity : Maintain visibility into all deployed AI Agent versions and application states across the environment through comprehensive asset inventory, helping detect potential supply chain vulnerabilities or unexpected execution activity, including RCE. Establish Visible Compliance Metrics : Help organizations prioritize risk and assess whether AI Agents comply with international risk-management and regulatory standards. 3. Flexible Deployment Non-disruptive cybersecurity Assessment : ThreatSonar Plus supports both online and offline deployment. Depending on environmental requirements, enterprises can conduct one-time cybersecurity scans without disrupting daily operations, quickly gaining visibility into AI Agent deployments and associated risks. An example of risk setting by the ThreatSonar Plus AI Agent for detection. Conclusion AI Agents are transforming enterprise workflows from “automation” to “autonomy.” AI is no longer merely an assistive tool; it is becoming an active “system participant” with real operational capabilities. As organizations benefit from the efficiency gains brought by AI, they must simultaneously evolve their cybercybersecurity mindset. Endpoint cybersecurity can no longer focus solely on monitoring files and processes—it must also understand and track AI behavior. By combining the command-level detection capabilities of ThreatSonar Plus with the real-time collaborative defense capabilities of ThreatSonar Anti-Ransomware , enterprises can embrace the AI wave while maintaining strong control over their digital environments. Want to find out how much Shadow AI or how many high-risk AI Agents may be operating within your enterprise environment? Contact TeamT5 and let us help you implement critical AI cybersecurity assessment and compliance measures.