Profet AI

emily

2026 Taiwan’s New AI and Cybersecurity Regulatory Landscape: A Governance Guide for Enterprises

By James Yang , Executive Assistant to the CEO, Profet AI

For the past few years, one of the most common questions businesses have asked about AI is: “What can AI do?”

As AI evolves from an assistive tool into an AI Agent, however, the questions businesses need to ask are becoming much broader:

What can AI see? What can it access? Whose authority can it act under? And if something goes wrong, can the organization detect it, contain it, and determine who is accountable?

This is why Taiwan’s evolving AI, cybersecurity, and data protection landscape deserves renewed attention in 2026.

In 2025, Taiwan amended the Cyber Security Management Act, which took effect on December 1, 2025. The Personal Data Protection Act was also amended in November 2025. In addition, the Artificial Intelligence Basic Act was passed by the Legislative Yuan on December 23, 2025, and promulgated on January 14, 2026.

These three laws address different areas. The Cyber Security Management Act focuses on information and communication systems and cybersecurity. The Personal Data Protection Act governs the collection, processing, use, and protection of personal data. The Artificial Intelligence Basic Act establishes fundamental principles for AI development and governance in Taiwan.

For enterprises, the key question is not simply “What new regulations have been introduced?”

It is also about understanding the broader governance direction that these regulations are collectively shaping.

As AI becomes embedded in core business processes, cybersecurity, data protection, and AI governance will increasingly converge within the same AI use cases.

To understand this regulatory shift, it is important to distinguish the roles of the three laws.

The Cyber Security Management Act primarily addresses how information and information and communication systems should be protected. Under the Act, cyber security involves preventing unauthorized access to, use of, control over, disclosure of, damage to, alteration of, or destruction of information and information and communication systems, while maintaining their confidentiality, integrity, and availability.

Its direct scope focuses on government agencies and specific non-government agencies, rather than applying uniformly to every private enterprise.

The Personal Data Protection Act addresses a different layer of risk.

When organizations collect, process, or use personal data, they need an appropriate legal basis and must adopt appropriate security and maintenance measures to protect individuals from improper harm.

The November 2025 amendment further strengthened the institutional framework for personal data protection, including supervisory and inspection mechanisms and the management of personal data security. However, as of the time of writing, the enforcement date of this amendment remains to be determined by the Executive Yuan.

The Artificial Intelligence Basic Act takes the discussion one step further by establishing fundamental principles for AI development and application in Taiwan.

Article 4 sets out seven principles for government promotion of AI:

  • Sustainable Development and Well-being
  • Human Autonomy
  • Privacy Protection and Data Governance
  • Cyber Security and Safety
  • Transparency and Explainability
  • Fairness and Non-discrimination
  • Accountability

Taken together, these three laws reveal a clear governance structure:

The Cyber Security Management Act addresses cybersecurity management for entities within its scope. The Personal Data Protection Act governs the collection, processing, use, and protection of personal data. The Artificial Intelligence Basic Act establishes fundamental principles for AI development and application, as well as the direction of future risk governance.

These three layers closely correspond to the challenges enterprises increasingly face as they move from traditional IT environments toward AI Agents.

Article content

Taiwan’s 2025 amendment to the Cyber Security Management Act further institutionalized cybersecurity requirements for entities within its scope through mechanisms including responsibility classification, cybersecurity maintenance plans, incident notification and response, audits, and outsourcing management.

In other words, organizations need to answer more than just:

“Do we have a firewall?”

They also need to ask whether they have a cybersecurity governance process that can operate continuously.

This becomes particularly important when a cybersecurity incident occurs. Relevant requirements emphasize incident notification, response, root-cause analysis, and subsequent improvement. Cybersecurity is therefore increasingly becoming part of enterprise operations and risk management—not simply a technical issue.

The arrival of AI Agents makes this challenge even more complex.

Consider a Prompt Injection attack that manipulates an Agent’s decision-making and causes it to deviate from its intended task.

The enterprise then needs to answer questions such as:

  • Why did the Agent have this permission?
  • Who authorized it?
  • What data did it actually access?
  • Were its actions fully logged?
  • Can the organization immediately restrict or stop it?

These questions are closely related to traditional cybersecurity governance.

The difference is that the enterprise now has another type of actor in its digital environment: an AI Agent capable of acting on behalf of people and systems.

As discussed in our previous article, supply chain security is another area that deserves renewed attention.

For entities covered by the Cyber Security Management Act, requirements related to outsourcing information and communication system construction, maintenance, and information and communication services include selecting appropriate contractors, requiring effective cybersecurity management mechanisms, and defining rights, obligations, and liability for breaches in written contracts.

This reflects a practical reality: an organization’s digital boundary no longer ends at its own data center or internal network.

Today, enterprises may simultaneously rely on cloud services, SaaS platforms, system integrators, third-party APIs, and AI model services.

Once AI Agents are introduced, a seemingly simple application may connect to enterprise knowledge bases, CRM systems, ERP platforms, APIs, and other tools—creating a new form of digital supply chain.

Organizations therefore need to look beyond a simple question such as:

“Does the vendor have a cybersecurity certification?”

They also need to ask:

  • What data can third parties access?
  • Which tools can an Agent invoke?
  • How are API tokens managed?
  • How long is data retained?
  • How are cybersecurity responsibilities divided when an incident occurs?

The growing focus on SBOM (Software Bill of Materials) can also be understood in this context.

SBOM is an important practical tool for improving transparency and management across the software supply chain. It is therefore more accurate to view SBOM as a tool for implementing and operationalizing governance, rather than describing it as a requirement that the Cyber Security Management Act explicitly imposes on every enterprise.

For enterprises, the more important direction is clear:

Supply chain security increasingly needs to be visible, controllable, and accountable.

This also means that IT outsourcing and AI service adoption should not be evaluated solely by procurement or IT teams. Cybersecurity, legal, data governance, and operational risk should increasingly be part of the same assessment.

If the Cyber Security Management Act encourages organizations to reconsider “Is our system secure?”, the Personal Data Protection Act raises another fundamental question:

“Do we actually have the ability to manage the personal data we hold?”

The November 11, 2025 amendment further strengthened the supervisory framework for personal data protection, including inspection and security maintenance mechanisms. The amendment also introduced and revised provisions concerning the supervision and management of personal data protection for non-government agencies. Its effective date, however, remains to be determined by the Executive Yuan.

This is particularly relevant to AI adoption.

In a traditional enterprise environment, a data flow might be relatively straightforward:

Employee → CRM → Customer Data

With AI, the flow may become:

Employee → AI Agent → Enterprise Knowledge Base → CRM API → AI Model → Response

As data moves across more systems, models, and tools, organizations need to ask more than whether a database is encrypted.

They need to understand:

  • Which data can enter an AI system?
  • Which data should never enter it?
  • What can third-party model providers access?
  • Can an Agent transfer data to another system?
  • Can these activities be traced and audited?

This is why AI Governance cannot be completely separated from Data Governance.

If enterprises want AI to become part of their core operations, data classification, access control, purpose limitation, retention, transmission, and access logging need to be considered from the beginning—not added as a privacy layer after the AI system goes live.

Among the three laws, the Artificial Intelligence Basic Act represents one of the most significant developments for Taiwan’s AI ecosystem in 2026.

The Act was passed by the Legislative Yuan on December 23, 2025, and promulgated on January 14, 2026. Its objectives include promoting human-centered AI research and development, fostering AI industry growth, building a secure AI application environment, protecting fundamental rights, and strengthening Taiwan’s international competitiveness.

Its significance lies in establishing a common set of governance principles for AI development and application in Taiwan.

Article 4 identifies seven principles, including human autonomy, privacy protection and data governance, cybersecurity and safety, transparency and explainability, fairness and non-discrimination, and accountability.

For example, the principle of Human Autonomy emphasizes human oversight.

Privacy Protection and Data Governance highlights data minimization and the prevention of data leakage.

Cyber Security and Safety calls for cybersecurity safeguards to ensure the robustness and security of AI systems.

And Accountability emphasizes corresponding internal governance and external social responsibility.

The direction is becoming even more concrete.

On July 7, 2026, Taiwan’s Ministry of Digital Affairs established the AI Risk Taxonomy and Assessment Framework, effective immediately. The framework is intended to support government agencies and sectoral authorities in identifying, assessing, and responding to AI risks, as well as developing risk-based management requirements.

For enterprises, this signals an important transition:

AI governance is gradually moving from shared principles toward more concrete, risk-based management approaches tailored to specific applications and industries.

This direction is especially important as more enterprises adopt AI Agents.

The greater the degree of autonomy an AI system has, the more mature the organization’s governance capabilities need to become.

Consider a company deploying a customer service Agent.

The Agent can access customer information in a CRM system, search the enterprise knowledge base, check order status, and even execute certain service processes through APIs.

From a product perspective, this is an AI application designed to improve customer service efficiency.

From a governance perspective, however, it involves three different layers of risk.

First, if the enterprise falls within the scope of the Cyber Security Management Act, issues such as accounts, permissions, system access, monitoring, incident response, and outsourcing may fall within applicable cybersecurity management requirements.

Second, the Personal Data Protection Act raises questions about whether customer information in the CRM can be processed in this way, whether the data is appropriately protected, and whether it is transferred to other services.

Third, the Artificial Intelligence Basic Act provides a broader governance direction, including data governance, cybersecurity measures, human oversight, and accountability.

From an enterprise governance perspective, these three laws can therefore be understood as three interconnected layers:

System Security Data Governance AI Governance

As AI Agents enter core business processes, mature organizations should consider these risks when designing the AI architecture itself—not after the system has already been deployed.

Article content

Faced with rapidly evolving regulations and AI technologies, enterprises do not necessarily need to build a highly complex governance framework from day one.

A more practical approach is to begin with eight fundamental questions:

Which are legal obligations, and which are regulatory guidance, industry standards, or internal governance requirements?

Which data and systems are the organization’s most important assets to protect?

Beyond employee accounts, which AI Agents can access these assets?

What data can each Agent read, modify, or act upon?

What data can the organization’s LLMs, cloud services, SaaS platforms, and APIs access or process?

Can the organization trace which data an Agent accessed, which tools it invoked, and which actions it performed?

When an Agent behaves unexpectedly, can the organization immediately restrict or revoke its permissions?

Are AI risks incorporated into enterprise governance, or do they remain technical documentation owned solely by the IT team?

Ultimately, these eight questions are asking whether an organization can truly:

See it. Control it. Trace it.

If the answer is no, successfully deploying AI does not necessarily mean the organization is ready for AI to scale into its core operations.

The value of regulation is not simply to help organizations avoid penalties.

More importantly, regulation encourages organizations to reconsider operational risks that may previously have been overlooked.

As enterprises increasingly depend on cloud services, APIs, AI models, and Agents, the relationships among data, systems, vendors, and people are becoming more complex.

This is particularly relevant to manufacturing.

AI Agents may eventually become part of processes across the manufacturing environment, where they interact with engineering data, equipment information, operational systems, and frontline know-how.

Once AI moves beyond providing recommendations toward executing tasks, enterprises need to address data governance, permissions, workflows, cybersecurity, and human oversight as part of the same operational architecture.

AI Governance should therefore become part of enterprise operational capability from the moment the AI architecture is designed.

Organizations need mechanisms covering the full lifecycle—from AI development and permission configuration to data access, tool usage, execution monitoring, incident response, and lifecycle management.

In 2026, the questions enterprises ask about AI are changing.

Yesterday, the question was:

“Can AI improve our efficiency?”

Today, it is increasingly becoming:

“Can AI safely enter our core operations?”

From the cybersecurity governance and outsourcing requirements under the Cyber Security Management Act, to data protection and security requirements under the Personal Data Protection Act, and the principles of human autonomy, data governance, cybersecurity, transparency, and accountability established by the Artificial Intelligence Basic Act, Taiwan’s enterprises are facing more than a series of individual regulatory updates.

A broader AI Governance environment is taking shape—one where cybersecurity, data protection, and AI governance increasingly intersect.

For enterprises, the real question may not be whether to add another compliance checklist.

It is whether they can build an operating model in which AI can be trusted, monitored, controlled, and held accountable.

As AI evolves from a tool into an actor within enterprise workflows, organizations need more than increasingly capable models.

They need more mature governance capabilities.

Compliance is not the destination of AI adoption. It is the starting point for building trusted AI operations.

2026 Taiwan’s New AI and Cybersecurity Regulatory Landscape: A Governance Guide for Enterprises Read More »

When the Security Boundary Extends Beyond a Single Enterprise: New Cybersecurity Challenges for the Semiconductor Supply Chain

By James Yang , Executive Assistant to the CEO, Profet AI

In our previous article, we explored the cybersecurity concerns surrounding the adoption of AI in the semiconductor industry. This time, we turn our attention to semiconductor supply chain security and examine why cybersecurity has become increasingly important to maintaining operational resilience and competitive advantage across a highly complex and interconnected supply chain.

Every stage of semiconductor manufacturing relies on technologies, equipment, data, and services provided by different organizations. This high degree of specialization enables the industry to maintain its enormous production scale and efficiency, but it also makes enterprise cybersecurity boundaries increasingly difficult to define.

In the past, enterprise cybersecurity largely focused on protecting an organization’s own networks, devices, and systems. Today, however, external suppliers may have remote access, third-party software may connect to internal systems, SaaS platforms may integrate with enterprise data through APIs, and suppliers themselves may depend on additional third parties.

As a result, the scope that organizations need to manage has extended far beyond their own IT environments.

If one node in this ecosystem is compromised, attackers may leverage existing trust relationships and access privileges to introduce risk into the enterprise—and potentially into other organizations as well.

For the semiconductor industry, supply chain security is therefore no longer simply an additional component of supplier management. It is increasingly becoming part of enterprise operational resilience.

This is not merely a hypothetical scenario.

In 2025, SecurityScorecard, a company specializing in cybersecurity risk assessment and supply chain security, surveyed nearly 550 cybersecurity leaders responsible for security strategy and management. More than 70% of respondents reported that their organizations had experienced at least one significant third-party cybersecurity incident in the previous year. The research also indicated that fewer than half of organizations were able to monitor the cybersecurity posture of more than 50% of their extended supply chain.

This highlights a challenge beyond the increasing number of third-party incidents: organizations may not have complete visibility into the risks across their own supply chains.

SecurityScorecard’s follow-up research in 2026 further found that 78% of respondents said their organization’s supply chain security programs covered less than half of their supplier ecosystem.

As the number of suppliers, services, and digital connections continues to grow, organizations’ ability to oversee these relationships may not expand at the same pace. The supply chain can therefore become a risk network that is increasingly difficult for an individual organization to fully control.

The npm supply chain attack in September 2025 provides one example.

The incident affected 18 widely used JavaScript packages. According to Unit 42, the threat intelligence team at Palo Alto Networks, these packages collectively accounted for more than 2.6 billion downloads per week. Attackers compromised the account of a package maintainer and injected malicious code into software components that had previously been trusted.

As a result, downstream users could potentially introduce compromised code into their own development or runtime environments simply by installing or updating the affected packages as part of their normal software workflows.

What makes this incident particularly significant is not only that third-party software was compromised, but that attackers exploited trust within the supply chain.

Downstream organizations did not directly interact with the attackers. They simply used software components they already trusted, yet that trust could potentially make them part of the attack chain.

A similar challenge can arise from integrations between SaaS platforms and third-party applications.

In November 2025, Salesforce identified suspicious activity associated with connections to the Gainsight application. The connections may have enabled unauthorized access to certain customer data. Salesforce subsequently revoked the application’s access and stated that there was no indication that the incident originated from a vulnerability in the Salesforce core platform itself.

The incident illustrates that supply chain security risks do not necessarily originate from a supplier’s own products or systems. They can also emerge from integrations, trust relationships, and access privileges between an enterprise and its third-party applications.

Attackers do not always need to directly breach an organization’s core platform. They may instead look for attack paths through third-party connections that already have legitimate access.

Together, these cases highlight a fundamental point:

Attackers do not necessarily need to break through an organization’s defenses directly. They may instead exploit trust, integrations, and access privileges between the organization and its third parties.

The challenge becomes even more complex in the semiconductor industry.

Semiconductor manufacturing relies heavily on equipment manufacturers, material suppliers, engineering service providers, and software vendors. Manufacturing equipment itself may involve operating systems, applications, network connectivity, remote maintenance, and data exchange.

In other words, a single piece of equipment may connect not just to one supplier, but to an entire ecosystem of organizations, systems, and services.

As discussed in our previous article, the semiconductor industry has already established SEMI E187, a cybersecurity standard for semiconductor manufacturing equipment.

In addition, in 2025, Taiwan’s Ministry of Digital Affairs worked with industry partners including SEMI, TSMC, and ASE to promote a verification framework based on the SEMI E187 semiconductor equipment cybersecurity standard. The framework was formally implemented in 2026, further establishing a third-party verification mechanism for equipment cybersecurity.

This represents a shift from security requirements defined by individual organizations toward industry-wide verification and trust mechanisms.

However, having a standard does not mean that the entire supply chain is secure.

SEMI E187 primarily focuses on semiconductor manufacturing equipment. In practice, organizations also face supply chain risks involving third-party services, remote maintenance, SaaS platforms, APIs, software components, supplier accounts, and suppliers’ own third-party dependencies.

Therefore, while SEMI E187 provides an important foundation for semiconductor supply chain security, it should not be treated as equivalent to comprehensive supply chain security governance.

A standard addresses “what should be done.”

Governance must go further and address “who is responsible, how it should be verified, how often it should be reviewed, and what should happen when the risk changes.”

Traditionally, organizations have approached supply chain security through supplier management, using questionnaires, audits, certifications, and contractual requirements to determine whether third parties meet defined security standards.

These mechanisms remain important.

However, as supply chains increasingly extend into suppliers’ own third parties, software components, SaaS platforms, APIs, and digital integrations, organizations need visibility beyond the security posture of individual suppliers.

They need to understand the critical dependencies, high-risk nodes, and pathways through which risk can propagate across organizations and systems.

Article content

This leads to one of the most practical challenges in supply chain security.

From the customer’s perspective, an organization may manage a large number of suppliers, each with different roles, access privileges, data exposure, and business criticality.

If every supplier is subject to the same cybersecurity requirements, organizations may need to invest significant resources in questionnaires, audits, and follow-up activities. At the same time, genuinely high-risk suppliers may become difficult to identify amid large volumes of lower-risk information.

The real question is therefore not simply:

“How many suppliers do we need to manage?”

It is:

“Which suppliers have the highest-risk connections to our critical assets, systems, equipment, and operations?”

For example, a supplier providing general materials and a maintenance provider with remote access to manufacturing equipment clearly present different levels and types of risk.

Likewise, an account with access only to general information should not necessarily be managed in the same way as a privileged account capable of accessing sensitive R&D data.

From the supplier’s perspective, cybersecurity requirements also have practical limitations.

For some small and medium-sized suppliers, additional security requirements may require significant investments in personnel, technology, and ongoing maintenance. When different customers impose different questionnaires, certifications, and security controls, suppliers may need to devote substantial resources to compliance activities without necessarily achieving a proportional reduction in actual risk.

Supply chain security therefore cannot simply mean that customers continuously add requirements and shift the responsibility and cost downstream to suppliers.

A more mature governance approach should align security requirements with the actual risks associated with each supplier, while establishing clear mechanisms for verification and continuous improvement.

One-time supplier questionnaires and audits still have value. They can help determine whether a supplier has established baseline controls and can serve as important references for procurement and contract management.

However, they cannot represent a supplier’s security posture several months—or several years—into the future.

Supplier systems may be upgraded, accounts and privileges may expand, remote maintenance practices may change, and the supplier’s own third parties may experience cybersecurity incidents.

The U.S. National Institute of Standards and Technology (NIST) describes Cybersecurity Supply Chain Risk Management (C-SCRM) as a systematic process for identifying, assessing, and mitigating cybersecurity risks throughout the supply chain.

A mature supply chain security program should therefore not rely solely on annual questionnaires or one-time audits. Instead, it should establish a continuous cycle:

Identify dependencies → Assess and control risk → Continuously monitor changes → Respond and improve

Audits answer the question:

“Are current requirements being met?”

Continuous risk management goes one step further:

“Has the risk changed?”

As supply chains become larger and more complex, it becomes increasingly impractical to impose exactly the same cybersecurity controls on every supplier.

Organizations should determine the appropriate level of security management based on factors such as the data, systems, equipment, access privileges, and business criticality associated with each supplier.

Lower-risk suppliers may be subject to simplified assessment and verification processes, while suppliers connected to core manufacturing environments, critical equipment, or sensitive data may require more comprehensive security assessments, access controls, technical verification, and continuous monitoring.

This does not mean lowering security requirements.

Rather, it means directing limited cybersecurity resources toward the areas where the potential impact is greatest.

For organizations, the key metric is not simply how many suppliers they manage, but whether they can clearly identify which suppliers could have a significant impact on critical operations if compromised.

If supply chain security is reduced to “customers telling suppliers what they must do,” it can easily become an exercise in transferring responsibility.

A customer adds cybersecurity requirements to a contract. The supplier submits documentation. Both parties complete an audit. The process appears to be finished—but the underlying risk may remain.

Effective supply chain governance requires customers, suppliers, and relevant service providers to establish clear responsibilities and coordinated response mechanisms.

Which risks are the supplier responsible for mitigating?

Which risks need to be managed jointly?

If a supplier experiences a cybersecurity incident, how quickly must it be reported?

Can the enterprise rapidly revoke access or isolate a connection?

These questions need to be defined before an incident occurs.

Supply chain security is therefore not about requiring every organization to achieve the same maximum security standard.

It is about establishing a governance model that can allocate responsibilities according to risk, continuously verify security conditions, and enable rapid coordination when incidents occur.

Article content

As supply chains extend to fourth parties, fifth parties, and beyond, it is not realistic for an organization to have direct visibility into every node.

The maturity of supply chain security should therefore not be measured simply by the number of suppliers an organization manages.

Instead, organizations need to determine whether they can identify critical dependencies, prioritize risks, and understand their exposure and potential impact even when they cannot directly control every part of the ecosystem.

For the semiconductor industry, what needs to be protected is not limited to a single piece of equipment, a single account, or a single supplier.

It also includes the process data, engineering know-how, intellectual property, equipment capabilities, and operational processes accumulated through cross-organizational collaboration.

These valuable assets are distributed across different organizations, systems, and services, and continuously move through equipment, software, people, and data.

The more complex the semiconductor supply chain becomes, the less feasible it is to rely solely on supplier management to ensure security.

The real challenge is to establish a governance framework that connects organizations, systems, and supply chain relationships while continuously identifying and managing risk.

In other words, organizations need to look beyond managing individual suppliers.

They need visibility into critical dependencies, high-risk nodes, access privileges, and data flows, while continuously adapting their security management as risks change.

When organizations move from “managing suppliers” to “governing the supply chain,” cybersecurity is no longer simply about passing requirements and responsibilities down to the next layer.

It becomes about building a governance capability that can identify critical dependencies, continuously monitor risk, and respond rapidly when an issue emerges at any point in the ecosystem.

For the semiconductor industry, the next layer of supply chain security is not simply about requiring every node to be secure.

It is about making the entire ecosystem more transparent, more verifiable, and more resilient.

When the Security Boundary Extends Beyond a Single Enterprise: New Cybersecurity Challenges for the Semiconductor Supply Chain Read More »

As Agentic AI Arrives, From SEMI E187 to Agentic Security: How Should the Semiconductor Industry Build Its Next Cybersecurity Defense?

By James Yang , Executive Assistant to the CEO, Profet AI

In our previous articles, we explored the risks enterprises may face when adopting AI Agents and the security governance mechanisms needed to address them.

This time, we take a closer look from the perspective of the semiconductor industry.

As enterprises increasingly adopt external AI services—and begin introducing AI Agents into manufacturing workflows—we need to ask whether the cybersecurity defenses we have built over the years are still sufficient.

For years, cybersecurity in the semiconductor industry has focused heavily on the security of equipment itself.

Does the equipment run a supported operating system? Are network communications secure? Are endpoints properly protected? Are security monitoring and logging capabilities in place?

Released in 2022, SEMI E187 was developed to address foundational cybersecurity requirements for semiconductor manufacturing equipment. It establishes security requirements across four key areas: Operating System, Network Security, Endpoint Protection, and Security Monitoring.

This provides a cybersecurity baseline that can be applied throughout the equipment lifecycle—from deployment and operation to ongoing maintenance.

But as AI Agents begin entering fabs and enterprise IT environments, the objects we need to secure are changing.

In the past, we managed equipment, servers, applications, and users. Now, we also have AI Agents that can understand tasks, access data, call tools, and potentially execute actions autonomously.

A recent report from Anthropic highlights another cybersecurity issue that semiconductor companies should pay attention to.

On September 10, Anthropic published its 154-page report, Detecting and Countering Misuse of AI, detailing how it detected and disrupted malicious activities involving Claude over the previous eight months, including cyberattacks, surveillance, fraud, weapons development, and unauthorized model distillation.

One of the most attention-grabbing findings was Anthropic’s disclosure that a user suspected of being connected to a Chinese military research institution had used Claude to develop software related to electronic warfare and air defense suppression, and had reportedly used Taiwan’s air defense facilities as simulated targets.

The headlines are certainly striking. But after reading the report, I found myself thinking about a question that is much more directly relevant to enterprises:

The ability to identify accounts, analyze usage patterns, connect multiple interactions, and ultimately develop threat intelligence indicates that AI platforms themselves need sophisticated capabilities for monitoring, detection, classification, and investigation.

There is a legitimate reason for this. Without such capabilities, AI platforms would have difficulty detecting and preventing fraud, cyberattacks, and other forms of abuse.

But the question changes when we look at this from an enterprise perspective.

In recent discussions with large semiconductor, AI server, and high-tech manufacturing companies, I have observed a common pattern: many enterprises have already begun adopting tools such as Claude Enterprise and ChatGPT Enterprise, giving engineers direct access to Frontier Models.

The productivity benefits are clear. But enterprises also need to understand:

  • What data are employees actually sending to AI?
  • Which conversations are retained, and for how long?
  • Which interactions may be subject to safety or abuse detection?
  • Under what circumstances could content be investigated further?
  • Can the enterprise control its retention policy?
  • Does the enterprise maintain its own audit logs?
  • Which process parameters, yield data, equipment information, and customer confidential information should never be sent directly to external AI services?

To be fair, Anthropic publicly states that inputs and outputs from its commercial products are not used to train models by default. Claude Enterprise also provides controls for data retention and auditing.

However, “not used for model training” is not the same as “never processed, retained, or subject to security detection.”

Anthropic’s documentation also explains that certain content associated with policy violations may be retained longer for safety investigations, while some Covered Models are subject to additional safety retention and review requirements.

This does not mean that every enterprise conversation is manually reviewed, nor does it justify concluding that enterprise secrets have been exposed. Actual data handling and retention depend on the specific product, enterprise agreement, and configuration.

But the report does highlight an important principle:

Enterprises should not only ask whether an AI provider uses their data to train models. They also need to understand who controls the data throughout the AI lifecycle, how it is processed, how long it is retained, and whether those activities can be audited.

As AI moves from answering questions to taking autonomous actions, these questions become even more important.

Historically, one of the major risks of external AI was that employees might voluntarily submit sensitive information to an AI service.

Agentic AI changes the equation.

AI Agents are no longer limited to answering questions. Through APIs, MCP, and other interfaces, they can access enterprise systems, retrieve data, and potentially execute real-world operations.

For example, an Agent responsible for process anomaly analysis may access MES, production history, equipment logs, and quality data at the same time.

Another Agent responsible for equipment maintenance could access machine status through connected tools and potentially recommend parameter adjustments.

These applications can significantly improve efficiency.

But they also introduce a new set of security questions:

Which Agent can access which data, call which tools, and execute which actions—and under what conditions?

This is one of the fundamental differences between Agent Security and traditional equipment security.

Within semiconductor manufacturing environments, SEMI E187 can establish a relatively clear equipment security governance chain through equipment procurement, supplier management, and ongoing maintenance requirements.

AI Agent adoption, however, may span IT, manufacturing, equipment engineering, quality, and multiple enterprise applications. Responsibility may not sit with a single organizational unit.

Enterprises therefore need to address at least three additional areas:

1. Dynamic Authorization Can permissions be dynamically adjusted based on the user, task, tools involved, and data sensitivity?

2. Agent Delegation When one Agent invokes another Agent, can permissions be properly constrained to prevent privilege escalation during delegation?

3. Identity Lifecycle Management When an Agent is created, activated, suspended, or revoked, can its credentials and access permissions be managed accordingly?

For equipment or human users, permissions may remain relatively stable over long periods.

For Agents, however, permissions need to evolve with the task and execution context.

This is why Agent Security cannot be solved simply by extending traditional account management.

Article content

SEMI E187 establishes four foundational cybersecurity areas for semiconductor equipment: Operating System, Network Security, Endpoint Protection, and Security Monitoring.

As AI Agents begin entering fabs and manufacturing workflows, we can borrow the layered security mindset behind E187 when thinking about how to build an Agent Security framework.

However, it is important to clarify that the following is a proposed extension of the governance concept inspired by E187—not a claim that SEMI E187 itself defines Agent Security requirements.

1. Identity & Runtime Security

Every Agent needs a clearly defined digital identity.

Who created it? Which user or business function does it represent? Where does it run? Who issues, rotates, and revokes its credentials?

When an Agent operates across multiple systems, enterprises must distinguish between the human user’s identity, the Agent’s identity, and the identity actually used to perform an operation.

An Agent should not be given a long-lived Service Account with broad access simply because it needs to perform one specific task.

2. Network & Tool Permission

When Agents interact with enterprise systems through APIs, MCP, or other interfaces, security is no longer just about whether the connection is encrypted.

The enterprise must also determine whether each tool invocation is appropriately authorized.

Which Agents can connect to MES? Which can query equipment data? Which can execute equipment-related operations?

RBAC and ABAC can provide the foundation for different levels of authorization, while access decisions can incorporate factors such as the user, task, equipment state, and operational risk.

For high-risk actions such as equipment parameter changes, organizations should consider additional controls such as Human-in-the-Loop approval, MFA, and short-lived authorization tokens.

3. Data Boundary & Isolation

Semiconductor manufacturers hold extensive process Know-how, equipment parameters, yield data, and customer-sensitive information.

During execution, an Agent may access these datasets while also generating conversation records, temporary files, working memory, and other intermediate artifacts.

Data governance therefore cannot stop at what a user initially submits to an AI model.

It must also address:

What did the Agent access? What did it generate? And what information could potentially be transmitted to an external ance therefore cannot stop at what a user initially submits to an AI model.

It must also address:

What did the Agent access? What did it generate? And what information could potentially be transmitted to an external model?model?

Clear data access and isolation boundaries are needed across departments, fabs, customers, and tasks.

This is why enterprises cannot rely solely on an external model provider’s data protection commitments. They also need a Data Boundary that they can control themselves.

4. Monitoring & Audit

Equipment security requires logs. Agents require them even more.

An Agent’s behavior is not necessarily a fixed sequence of actions. Given different data and contexts, the same task may lead an Agent to select different tools and execution paths.

Enterprises therefore need visibility into:

Who initiated the task? Which Agent executed it? What data did the Agent access? Which tools did it call? What intermediate artifacts or working memory were generated? Which actions were authorized? Which were blocked by guardrails? If something goes wrong, can the activity be traced, permissions restricted, or execution stopped immediately?

These records need to be integrated, retained, and transformed into auditable reports.

Only then can Agent activity become a manageable enterprise security capability.

Article content

Let’s return to Anthropic’s report.

It demonstrates that major AI platforms already possess sophisticated capabilities for abuse detection and threat investigation. More importantly, it gives enterprises an opportunity to rethink the security capabilities they need to establish themselves.

For large semiconductor and high-tech manufacturing companies, an AI Agent platform needs to govern more than the model itself.

From models and data to tools and actual operations, enterprises need an AI Governance Layer that they control.

This governance layer should cover:

Identity → RBAC / ABAC → Guardrails → Tool Permission → Data Boundary → Audit Log → Retention → Reporting

Together, these capabilities form a closed-loop AI security governance framework.

When these controls can be managed consistently, enterprises can maintain the same security requirements across different Frontier Models, different Agents, and different enterprise systems.

This does not replace SEMI E187.

Nor does it mean that existing equipment security is no longer sufficient.

Rather, as AI Agents begin to access enterprise data and execute actions, a security architecture historically centered on equipment, networks, and human users needs to extend toward Agent identity, behavior, and authorization.

The future question for semiconductor security is no longer only:

Which equipment is allowed into the fab?

It is also:

Which Agent is allowed in? What can it see? What can it do? And if it makes a mistake, can we stop it before it creates real-world impact?

AI should—and will—be widely adopted.

But enterprises cannot allow the convenience of AI to mean outsourcing their data boundaries, permissions, and governance along with it.

After reading Anthropic’s 154-page report, I believe the question enterprises should be asking is not simply whether external AI platforms have adequate security protections.

It is this:

When thousands of employees use AI every day and hundreds of Agents begin connecting to enterprise systems, can the enterprise still see what data is leaving its boundaries—and what actions are being executed in the name of AI?

That is the question the semiconductor industry needs to start answering as it moves from Equipment Security to Agent Security.

As Agentic AI Arrives, From SEMI E187 to Agentic Security: How Should the Semiconductor Industry Build Its Next Cybersecurity Defense? Read More »

Putting Agent Security into Practice: How Do We Verify an Agent’s Identity—and Contain It When Things Go Wrong?

By James Yang, Executive Assistant to the CEO, Profet AI

How Taisys, Profet AI, and Zentera illustrate a complete Agentic Zero Trust defense—from hardware-based identity and platform governance to network isolation

In the previous three articles, we explored the threats facing AI Agents, the principles of Zero Trust architecture, and the control processes required to move from security design to enterprise deployment.

Once AI Agents actually enter the enterprise environment, however, a more practical question emerges: How do we turn these security principles into controls that can actually be enforced?

As AI Agents evolve from answering questions to accessing enterprise data, calling MCPs and tools, interacting with backend systems, and even executing tasks with direct business impact, Agent Security is no longer limited to model security. It has become a comprehensive control chain spanning identity, authorization, platform governance, endpoints, and networks.

Architecturally, this defense can be viewed across three distinct layers.

The first asks: Who is actually authorizing this Agent?

The second asks: What is this Agent allowed to do?

The third asks: If the Agent goes out of control, where can it actually reach?

The technologies represented by Taisys, Profet AI, and Zentera map naturally to these three security boundaries.

Traditional enterprise IAM primarily manages the identity of people.

A user authenticates through a username and password, often combined with multi-factor authentication, and is then granted access based on the assigned permissions.

Agentic AI changes this fundamental assumption.When an Agent can continuously act on behalf of a user, enterprises need to verify not only who the user is, but also which Agent is acting, and whether a high-risk operation has actually received human authorization at the moment it matters.

Taisys has extensive experience in SIM, eSIM, and telecommunications identity technologies. One of its key strengths is leveraging the secure hardware embedded within SIM-based technologies to establish a trusted identity foundation.

A SIM card provides characteristics similar to a Secure Element. Sensitive identity credentials and cryptographic keys can be protected within hardware, while cryptographic operations can be performed within the secure environment rather than exposing critical credentials directly to the operating system, browser, or Agent runtime.

This model has been applied for years across telecommunications and financial services. The direction represented by dMID is to extend this established hardware-based trust model toward AI Agent scenarios.

Today, many Agent systems implement Human-in-the-Loop through instructions such as:Before deleting data, making a payment, or modifying a critical setting, ask the user for confirmation.

The problem is that when this rule exists only inside the Agent’s context, the security control remains within the same trust boundary as the Agent’s reasoning.

If the context is compressed, instructions conflict, or the Agent misinterprets the task, the intended approval process may fail.

There have already been notable Agent-related incidents in 2026, including an email Agent deleting large volumes of emails without completing the expected confirmation process, and a coding Agent accidentally deleting production data while troubleshooting an environment issue.

The common factor was not necessarily an external attacker stealing credentials.

Rather, the Agent already had legitimate access—and then used that access incorrectly.

This leads to an important security principle:

Final authorization for high-risk operations should not exist solely inside the Prompt.

A more robust architecture is to allow the Agent to propose an action, while requiring an independent trusted mechanism outside the Agent runtime to re-authenticate the user and verify the authorization before the action is actually executed.

A SIM-based Secure Element can provide a potential Hardware Root of Trust for this model.

In other words: The Agent can determine what needs to be done. It should not be able to prove that the human has approved it.

This separates Human Approval from the Agent’s reasoning process and creates an independent hardware-backed security boundary.

Even if the Agent runtime malfunctions—or is compromised—a high-risk action must still cross a separate trusted authentication boundary before it can be executed.

But verifying who approved an action is only the first step.

Once identity has been established, enterprises must also determine what that user is actually allowed to do through a particular Agent and within a particular context.That is where Agent Governance comes in

Article content

In an enterprise Agent architecture, Identity and Authorization must be treated as separate concepts.

Authenticating a user does not mean that the user should be able to access every enterprise resource through every Agent.

Likewise, an Agent with a legitimate identity should not automatically be allowed to access every Knowledge source, Model, MCP, Tool, or enterprise backend system.

This is the core idea behind how Profet AI positions AI Studio as an enterprise AI Agent operation platform.

AI Studio is not simply managing individual Agents. It manages the governance relationships among the different components involved in Agent execution, including Users, Agents, Workflows, Knowledge, Models, AutoML, MCPs, Tools, and Enterprise Systems.

Consider an enterprise FAE Twin.

It may need to search historical technical documents, analyze customer issues, call an AutoML model to analyze manufacturing data, and interact with enterprise systems such as Jira, PLM, or other backend platforms through MCP.

The real governance questions are therefore no longer limited to:

“Can the FAE Twin log in?”

Instead, enterprises need to ask:

  • Which customer’s data can it access?
  • Which information is read-only?
  • Which systems can it modify?
  • Which Tool Calls can be executed automatically?
  • Which actions require human approval?

These decisions should not be left to the LLM itself.

AI Studio uses mechanisms such as ACLs, Guardrails, Workflows, Knowledge Governance, Tool/MCP Management, and Audit to place enterprise policies outside the Agent’s reasoning process.

The Agent can understand the request, analyze information, and plan a task.

But whether a resource can be accessed, whether a tool can be called, and whether an action requires approval should be determined by an independent policy layer.

A complete Agent control architecture should therefore separate at least three functions:

Reasoning is handled by the Agent.

Policy Enforcement is handled by the governance platform.

Human Approval provides the final authorization for high-risk or irreversible actions.

If all three are embedded in the same Prompt, the Agent effectively becomes the applicant, reviewer, and executor at the same time.

That is fundamentally inconsistent with the principles of Zero Trust.

As AI moves from providing information to taking action, security controls must therefore expand beyond the model layer to cover the entire Agent Execution Chain.

Yet even with trusted identity and platform-level authorization in place, one critical question remains:

What happens if the Agent itself is compromised by Prompt Injection, a malicious MCP, or a third-party tool?

And what happens when an enterprise has Shadow AI that has never been formally registered or governed?

Application-level policies alone may not be enough.Security must extend further down into the infrastructure and network layers.

Article content

Zentera Systems has long focused on Zero Trust Network Security. Its CoIP Zero Trust Overlay uses software-defined isolation and micro-segmentation to establish security boundaries around enterprise applications, data, and workloads.

The underlying principle is straightforward:

Even if an attacker compromises one endpoint or obtains valid credentials, they should not automatically be able to move laterally across the enterprise network.

Ensage AI extends this Zero Trust approach to AI Agents.

The collaboration between Profet AI and Zentera brings Agent Governance and underlying Zero Trust network enforcement together.

AI Studio manages the logical authorization relationships between Agents and resources such as Knowledge, Models, MCPs, Tools, and enterprise applications.

Ensage AI extends this control into the Endpoint, Session, and Network layers.

Through zLink, Ensage AI can perform Agent Discovery based on Process Fingerprints and Network Behavior, helping enterprises identify Coding Agents, MCP Servers, and third-party tools based on their actual presence and behavior rather than relying solely on registration.

This addresses the growing challenge of Shadow AI visibility.

At the same time, AI Session Controller provides visibility and control over Prompts, Responses, and Tool Calls, while an intermediary layer can help reduce sensitive credential exposure and limit the potential for compromised credentials to be reused for lateral movement.

Combined with ecosystem risk tracking from Zentera Labs and Virtual Chamber technology, different Agents and workloads can be further isolated within independent network environments.

Consider a simple example.

An Agent working on Project A should not be able to access resources belonging to Project B.

This should not depend solely on a platform policy saying : “Access denied.”

At the network layer, Project B’s services should simply not be reachable by the Agent in the first place.

Likewise, a development Agent compromised through Prompt Injection should not suddenly gain network reachability to a financial system.

A Coding Agent may have permission to interact with a Git repository, but that does not mean it should be able to establish a direct connection to a production database.

This is a critical concept in Zero Trust : Blast Radius Control

Traditional permission design typically starts with the question:

“What permissions does this Agent need to complete its job?”

Zero Trust requires an additional question:

“If this Agent completely loses control one second from now, how much of the enterprise are we willing to let it affect?”

The first question ensures that the Agent can complete its task.

The second determines the maximum potential damage if something goes wrong.

For AI Agents, this distinction is becoming increasingly important. Agents can execute APIs, Commands, and Tool Calls far faster than a human security team can respond. Relying solely on monitoring and alerts after an incident has already occurred may simply be too late.

Therefore, some security controls must exist at an infrastructure layer that the Agent itself cannot bypass.

Article content

When Taisys, Profet AI, and Zentera are viewed together, it becomes clear that each addresses a different security boundary.

Taisys dMID uses a SIM-based Secure Element as a hardware trust foundation to protect identity credentials and cryptographic mechanisms, while providing an independent trusted authentication mechanism before critical actions are executed.

The core question this layer answers is:Who actually approved this action?

Profet AI AI Studio governs the access relationships among Users, Agents, Workflows, Knowledge, Models, AutoML, MCPs, Tools, and Enterprise Systems.

Through mechanisms such as ACLs, Guardrails, Workflows, Approval, and Audit, it determines what an Agent is allowed to do within a specific context.

The core question is:What is this Agent allowed to do right now?

Zentera Ensage AI extends control from Agent Discovery and Endpoint Visibility to AI Session Control, Micro-segmentation, and Virtual Chamber-based isolation.

It limits the actual network and system environments that an Agent can reach.

The core question is:Even if the Agent is compromised or goes out of control, where can it actually go?

These three security questions are closely related, but they cannot replace one another.

Identity verification cannot by itself prevent a legitimately authorized Agent from moving laterally across the enterprise network.

Platform governance alone cannot guarantee visibility into every unmanaged Shadow Agent running on endpoints.

Network isolation cannot determine whether a high-risk business action has actually received human approval.

A true Agentic Zero Trust architecture therefore requires these three control layers to remain relatively independent while forming a consistent Trust Chain.

Article content

Agentic AI security is unlikely to be solved by a single product.

Agent Identity, Human Approval, Agent Governance, MCP and Tool Management, Endpoint Visibility, Session Control, and Network Isolation represent different areas of expertise.

A more realistic enterprise architecture is therefore not to expect a single Agent platform to recreate every security capability.

Instead, different security layers should each enforce the controls they are best positioned to provide, while connecting through consistent policies, identity, and auditing mechanisms.

Within this architecture :

Taisys provides a Hardware Root of Trust through SIM-based secure hardware.

Profet AI AI Studio serves as the Policy Control Plane for Agent Governance.

Zentera provides Endpoint, Session, and Network Enforcement.

Together, these technologies are not simply three independent products.They can be viewed as components of a broader trust chain spanning identity, authorization, execution, and isolation.For enterprises, this also changes how Agentic AI platforms should be evaluated.

The key questions should no longer be limited to :

  • How capable is the model?
  • How many Agents can the platform support?
  • How many MCPs and tools can it connect?

Instead, enterprises should be asking :

Who initiated the action?

Which Agent executed it?

What policy granted the Agent permission?

Who approved the high-risk action?

Which resources were actually accessed?

And if something goes wrong, can the action be immediately revoked, isolated, and fully traced?

As AI Agents move from information assistants into core enterprise workflows, the focus of security must evolve beyond “preventing the model from giving the wrong answer.”

The real challenge is to ensure that every autonomous action taken by an Agent exists within a trust boundary that is verifiable, authorized, constrained, and auditable.

That is the problem Agentic Zero Trust ultimately needs to solve.

Putting Agent Security into Practice: How Do We Verify an Agent’s Identity—and Contain It When Things Go Wrong? Read More »

From Threats to Framework: Seven Capabilities in the Agentic AI Era, Where Does Your Enterprise Stand?

By James Yang, Executive Assistant to the CEO, Profet AI

After finishing the previous article, if you are calculating which tool to buy to address each type of threat, pause for a moment. The mindset of patching vulnerabilities one by one could barely keep pace back in the era of human attackers. Now that attackers are also using AI agents, new attack techniques are emerging faster than security teams can patch them. This article is about switching to a completely different defense logic.

Zero Trust as a concept traces back to 1994, when Stephen Paul Marsh first systematically proposed it in his doctoral dissertation at the University of Stirling. NIST formally established concrete guidance with SP 800-207 in 2020. In its report “Zero Trust for AI Agents,” Anthropic applies this framework to the entirely new attack surface of agentic AI, proposing seven capability domains detailed down to an executable level.

The core logic of Zero Trust, in one sentence: trust nothing and no one, verify every access request, and assume a breach has already occurred. Expanded, this becomes three tenets:

•       Never trust, always verify: Every access request must be authenticated and authorized, regardless of whether it originates from inside the corporate network.

•       Assume breach: Systems should be designed on the assumption that intrusion will eventually happen. The focus is not on preventing intrusion, but on limiting the damage an attacker can do once inside.

•       Least privilege: Grant only the minimum access needed to complete a task, so that any single breach is contained to a small blast radius.

More important than the principles themselves is a litmus test: does this control make an attack impossible, or does it merely make the attack more inconvenient? Defenses that rely on “adding friction”, extra jump boxes, rate limiting, non-standard ports, lose most of their effectiveness against automated attackers whose cost per attempt approaches zero. Controls that actually hold up share a common trait: hardware-bound credentials, tokens that expire, cryptographic identities, and network paths that simply do not exist. This litmus test runs through all seven capability domains described below.

Article content

Anthropic breaks each capability domain into three tiers: Foundation, Enterprise, and Advanced. Enterprise builds on Foundation, and Advanced builds on Enterprise, each tier reinforcing existing controls layer by layer rather than starting over. Foundation is the entry point into Zero Trust. Enterprise is the target level most organizations of a certain size should aim for. Advanced is designed for heavily regulated industries, national security applications, or deployments where the consequences of failure are severe.

Worth noting: because of the compressed attack speed discussed in the previous article, the bar for Foundation itself has been raised. Practices once considered a reasonable starting point, such as managing risk through API key rotation, no longer qualify. Short-lived tokens, cryptographically rooted identities, identity-based isolation, and automated frontline alert triage are now baseline requirements, not advanced options.

Article content

Identity and authentication form the foundation underlying all other capabilities. Without a verifiable identity, access control cannot be enforced, and behavior cannot be attributed to a specific agent. The path runs from cryptographically rooted persistent identity, to certificate-based authentication paired with lifecycle management, and ultimately to hardware-bound identity, where credentials are stored in a hardware security module and integrity is verified through remote attestation.

Once identity is established, an agent must prove its identity again every time it accesses a service. This is service authentication. The path runs from short-lived tokens measured in minutes, to mutual TLS with certificate pinning, and ultimately to hardware-bound credentials. Managing risk through API key rotation is no longer sufficient, because credentials that can be extracted from a codebase do not become meaningfully harder to steal no matter how often they are rotated.

An agent with legitimate permissions has a genuine identity. The problem is what it does once authorized. The path runs from role-based access control with default deny, to attribute-based access control that factors in context such as time, location, and risk score, and ultimately to real-time authorization that is re-evaluated for every single action. OWASP’s concept of Least Agency extends least privilege to cover what each tool is allowed to do, how often, and where, for example limiting a database tool to read-only queries.

It is worth being precise about the difference between RBAC and ABAC. RBAC (role-based access control) only looks at what role an agent holds. Once a role is assigned, its scope of permissions is fixed, and every agent under that role has identical access regardless of context. ABAC (attribute-based access control) goes further and factors in the context of the request itself, such as time, location, data sensitivity, and risk score. Two agents may both be customer service agents, but one querying general customer data during normal hours will be allowed through, while the same agent pulling large volumes of highly sensitive data late at night may be denied or asked for additional verification. Moving from RBAC to ABAC is essentially upgrading from a static judgment of “who you are” to a dynamic judgment of “what you are doing right now and how risky it is.” This is the single most important upgrade in moving from Foundation to Enterprise.

While the permission model decides whether access should be granted at all, the scope of that access determines when and for how long it can be exercised. Static permissions, once granted, persist indefinitely, remaining active even after the task is finished and creating long-term exposure. Dynamic permissions are elevated only for the duration of task execution and are withdrawn immediately afterward. The strictest form, just-in-time and just-enough access (JIT/JEA), compresses even the act of elevation to nearly the moment of use, granting access only at the instant it is actually needed and automatically revoking it on timeout or task completion, leaving no residual access behind. Last is the boundary of resources: identity-based isolation as the primary control, supplemented by network segmentation, progressing to containerized sandboxing, and ultimately to hardware isolation. For agents that process untrusted input, sandboxed execution is standard practice, not optional.

Access control determines what gets blocked. Observability determines what actually happened. Logging progresses from complete records with timestamps, to tamper-evident audit trails, and ultimately to real-time streaming into a centralized monitoring platform. But before chasing the specifications of all three tiers, it is worth first establishing measurement for two metrics: dwell time (the interval between an anomaly occurring and a human noticing it) and coverage (the proportion of alerts that are actually investigated). These are the two metrics where AI automation delivers the greatest leverage.

Traceability is a separate dimension: generating a unique identifier for each request that follows it through every subsequent action, progressing to distributed tracing across multi-agent workflows, and ultimately to a complete provenance chain that supports replaying the full decision history. Logging answers “what happened.” Traceability answers “why it happened.”

Behavioral monitoring has to judge whether a given behavior is normal or suspicious, which requires an established baseline before anomalies can be identified. Baselines progress from being manually defined, to automatically learned statistical baselines, and ultimately to baselines that continuously update and can detect slow, gradual drift. This tier specifically addresses techniques such as long-term memory drift. Anomaly detection progresses from threshold-based alerts, to statistical methods, and ultimately to machine learning models that incorporate contextual information.

Once an anomaly is detected, response speed is what actually determines the extent of the damage. The division of labor is clear: automation handles the paperwork, logging, evidence collection, and drafting initial reports, while decisions about containment, disclosure, and customer communication remain with humans. Response capability itself also progresses through three tiers: alerts routed to the security team with a triage agent producing an initial assessment, progressing to automated execution of containment actions, and ultimately to agentic SOAR capable of coordinating containment across multiple systems. What automation should do is free up human response time, not replace final judgment.

This domain is the direct countermeasure to indirect prompt injection. On the input side, the path runs from basic format validation, to content filtering that screens out encoded payloads and suspicious instructions, and ultimately to multi-layer validation combining AI classifiers with content boundary markers. SQL injection has well-defined attack patterns that can be defended against, but agent input is free-form and unpredictable, so rule-based defense alone is not enough.

The output side guards against data exfiltration: progressing from scanning for sensitive patterns, to analyzing the semantics of output to identify social engineering risk, and ultimately to requiring human approval for high-risk actions. Input validation prevents malicious manipulation. Output control prevents data leakage. Both are essential.

When every other defense fails and a breach actually occurs, attackers often turn to tampering with the agent’s own configuration. The path runs from version control, to cryptographic signing, and ultimately to immutable infrastructure, where configurations are never modified but wholesale replaced, with image integrity verified before execution.

One practice runs counter to intuition: for components where risk is acceptable, enabling automatic updates is actually safer than insisting on manual approval, because the delay inherent in manual approval is itself a risk. The solution is to let automatic updates coexist with pre-deployment signature verification: signed updates flow through automatically, while unsigned changes are rejected outright. Recovery follows a similar path, from documented rollback procedures, to automated rollback, and ultimately to self-healing systems. No amount of automation can save a situation that cannot reliably roll back to a known good state.

Technical controls can only enforce what governance policy defines. The path runs from a documented acceptable use policy, to a cross-functional governance framework that includes security, legal, and business stakeholders, and ultimately to embedding policy checks into the deployment pipeline itself, automatically detecting violations. Shadow AI, meaning employees using LLM tools on their own without approval, bypasses every one of the controls described above. This is the reality governance has to confront.

Having gone through the seven capability domains, you do not need to redesign your defenses for every new attack technique that emerges. What you need to do is check where your enterprise’s existing agent deployments stand across these seven areas. Identity makes attribution and access control possible. Observability reveals what actually happened. Behavioral monitoring detects anomalies. Input and output controls intercept attacks at the boundary. Integrity makes recovery possible. Skip any one of these areas, and that is the gap attackers will come through.

The next article moves from framework to implementation: a concrete operational workflow, and how defense teams can accelerate their security operations to keep pace now that attack speed has compressed to a matter of hours.

From Threats to Framework: Seven Capabilities in the Agentic AI Era, Where Does Your Enterprise Stand? Read More »

When Attackers Also Wield AI Agents, the Ground Beneath Enterprise Defense Starts to Shift

By James Yang, Executive Assistant to the CEO, Profet AI

AI is destabilizing the balance of speed between attackers and defenders, and most enterprises have yet to grasp this. Palo Alto Networks Unit 42’s 2026 Global Incident Response Report, drawing on more than 750 major security incidents across over 50 countries, found that the fastest 25% of attacks now take an average of just 72 minutes from initial access to full data exfiltration, four times faster than the previous year. The same report found that attackers begin automated scanning an average of 15 minutes after a vulnerability is disclosed.

This velocity is not an isolated finding. Verizon’s 2026 Data Breach Investigations Report found that nearly one third of data breaches now originate from software vulnerabilities, surpassing stolen credentials as the leading initial access vector. On the defense side, the World Economic Forum and KPMG’s Empowering Defenders report offers a telling counterpoint: organizations with mature AI defense capabilities cut breach detection and response cycles by roughly 80 days and reduce average breach costs by up to 1.9 million US dollars.

Seventy two minutes. Fifteen minutes. This is the attacker’s timetable today. Can your security team keep pace?

Enterprises are deploying AI agents at an equally striking pace. BCG’s 2026 AI Radar found that 90% of CEOs worldwide expect AI agents to deliver measurable ROI this year. Gartner forecasts that by the end of 2026, 40% of enterprise applications will incorporate task specific AI agents, up from under 5% in early 2025.

Taken together, these numbers point to the same conclusion: AI does not take sides. It accelerates both defenders and attackers at once. The only question is who recognizes that the rules have changed first.

Conversations about AI security and governance have traditionally centered on the model itself: Is it safe? Will our data be used for training? What does the privacy policy say? These are valid questions, but they belong to the old battlefield. What deserves real attention in 2026 is something else entirely: as enterprises push AI agents to the front lines, attackers are arming themselves with the same technology, often well ahead of the enterprises they target. Governance itself has quietly become a competitive advantage.

Article content

Traditional security logic is sound: verify identity, confirm authorization, allow actions within scope. That logic has protected enterprises for two decades. But an AI agent operating with legitimate, properly issued credentials has not broken through any door it was not supposed to open. The problem lies in what it does once it has been authorized.

This is the most fundamental way AI agents differ from every prior category of security incident. Traditional access control governs who gets through the door, but an agent capable of reading databases, sending emails, and querying industrial control systems can be manipulated into taking an action that falls within its permissions yet was never meant to happen, and no access control can stop it, because from the system’s perspective, every credential is valid, every process is normal, and no rule has been broken.

The five threats unfolding right now all target the same blind spot: what is actually hiding behind operations that look entirely legitimate.

Article content

This is the blind spot most easily overlooked in this era. A conventional question answering AI only responds to input typed directly by a user. An AI agent is different: it browses the web, reads documents, and queries databases on its own initiative, gathering information and taking action autonomously. This capability opens an entirely new attack surface. An attacker only needs to embed a hidden instruction for the agent inside a web page, an email attachment, or a shared document. The root cause is a fundamental technical limitation: Microsoft Research has confirmed that large language models cannot reliably distinguish between text that is merely reference material and text that constitutes an instruction to be executed. The user never sees the malicious instruction, yet the agent treats it as a legitimate request and carries it out. When an agent is granted access to external files or industrial databases and integrated with MES, ERP, or OT control systems, the consequences escalate sharply. Research published in the journal Information (MDPI) in January 2026 found that in industrial control settings, the potential impact of this class of attack extends well beyond data exfiltration into direct manipulation of production systems. A Center for Internet Security report from April 2026 recorded a 340% year over year increase in related attacks, and the OWASP GenAI Security Project ranks prompt injection as the top threat to agentic AI.

Every tool an agent can call opens another door an attacker might exploit. An attacker can tamper with a tool’s description or interface, tricking the agent into trusting a compromised tool and executing an unauthorized action, a technique known as tool poisoning. Alternatively, an attacker can leave a tool clean for a period of time, then swap in a malicious version once the agent has come to trust it, a technique known as a rug pull. Anthropic has documented a real world case: the first malicious MCP server ever discovered in the wild posed as a legitimate email service while secretly copying every message the user sent. Harder still to detect is tool chaining. Viewed individually, an internal CRM lookup tool and an external email tool are each perfectly safe, but an attacker can induce the agent to chain them together, pulling customer records with the CRM tool and exfiltrating them through the email tool. The entire sequence runs on legitimate credentials through normal program execution, so conventional malware monitoring detects nothing at all.

Beyond tampering with the tools themselves, an agent’s automated nature can also be exploited directly. A human operator grows fatigued with repetitive tasks and pauses to check in, but an agent does not. As long as its logic concludes that one more query is warranted, it will keep triggering the next call. An attacker needs only to trap the agent in a loop, repeatedly invoking a costly external API, to cause a service outage or a runaway cloud bill within minutes. The entire attack is self triggered by the agent, requires no human intervention, and does not need to breach any system at all. It is often noticed only after the bill or the service alert arrives, a textbook case of legitimate functionality being overused rather than a conventional intrusion.

A common failure in multi agent systems is unscoped privilege inheritance: a high privilege manager agent delegates a task to a worker agent without applying least privilege scoping, passing along its full access rights unchanged. A related variant is the confused deputy problem: a lower privilege agent forwards an instruction that appears legitimate to a higher privilege agent, which executes it without verifying the original user’s intent. Risk can also hide in memory. If an agent caches credentials or keys from a prior session for reuse without proper memory segmentation, an attacker can prompt it to retrieve those residual credentials and carry out actions its own credentials would never permit, effectively escalating privilege across session boundaries. When multiple agents routinely trust and delegate to one another, these gaps widen further, and reconstructing which delegation caused the breach often only happens after the fact.

Agentic systems differ from traditional software in that they frequently assemble capabilities at runtime, loading external tools and switching between models or personas on the fly, which pushes the attack surface beyond what traditional software supply chain analysis can handle. The risk operates on two layers. One lies in the model itself: if training data or the fine tuning process is tampered with, a backdoor can survive the entire deployment lifecycle. Anthropic’s own research found that injecting just 250 malicious documents is enough to plant a backdoor in models ranging from 600 million to 13 billion parameters, and that backdoor persists even after subsequent safety training, including supervised fine tuning and RLHF. The other layer lies in tools and frameworks: the PyTorch dependency confusion attack showed how malicious packages can exfiltrate SSH keys during installation, and security researchers have identified roughly 100 malicious AI models on major platforms, some of which open a reverse shell the moment they are loaded. A model’s provenance, where its training data came from and whose hands it passed through, is itself an attack surface that warrants scrutiny.

Agents retain records of past interactions to become more useful, but without proper isolation, a single successful injection can poison every future session, leaving the agent to carry out an attacker’s planted instructions indefinitely. Contamination can arrive through more than one channel. If an agent relies on a vector database for retrieval (RAG), poisoned sources, direct uploads, or an overly trusted data pipeline can all introduce malicious data into that database, and anything the agent later retrieves from it can produce false answers or execute a planted payload. In multi tenant environments, contamination can also spread through shared context, passing from one user’s session into another’s.

These poisoning techniques can still be traced back to a single identifiable event. Long term memory drift is a harder to detect variant. It does not rely on one successful injection but instead works through summarization or feedback from other agents, gradually and quietly distorting the knowledge or goal weighting an agent has stored. Because each individual change looks harmless in isolation, perhaps a slight shift in how a summary is worded or a minor adjustment in weighting, conventional anomaly detection struggles to flag any single instance. It is typically only noticed once the agent’s behavior has visibly drifted from its original scope, and by then the drift may have been accumulating for a long time.

Having read through these five threats, it is easy to fall into a particular kind of anxiety: does every new attack technique require a new corresponding defense? Follow that logic and you will spend forever chasing attackers, because new techniques keep emerging faster than any team can patch for them. The real question is different: is there a set of principles that holds regardless of what new technique appears next, so that as long as the principles are upheld, the defense holds too.

If a service account’s access were scoped down from the outset to the minimum dataset its task actually requires, then even if it were manipulated into an anomalous action, the blast radius would already be confined to a small perimeter. The breach happens, but the damage is contained. That is an achievable defensive posture: not preventing every breach, but ensuring a breach never becomes the start of a disaster.

In the next installment, we break down a systematic framework for putting this into practice: eight capability domains across three tiers of maturity, showing exactly what a defense architecture built to withstand the agentic AI era looks like.

When Attackers Also Wield AI Agents, the Ground Beneath Enterprise Defense Starts to Shift Read More »