Introduction: The core of AI agent governance is shifting from "content review" to "action control"

Enterprise use of artificial intelligence is moving from the stage of "letting AI generate content" into the stage of "letting AI complete tasks." In the past, generative AI mainly handled work such as information retrieval, text drafting, meeting minutes, and coding assistance, and the risks enterprises focused on were mainly whether the output content was accurate, whether it infringed rights, whether it leaked input information, and whether it contained bias or inappropriate expression.

But the risk structure of AI agents has already changed. An agent does not merely answer questions; it may autonomously plan steps according to goals, call tools, access databases, modify system records, send emails, submit forms, execute code, trigger payments, or interact with third-party systems. In other words, the risk of AI is no longer only that it "says something wrong," but that it may "do something wrong." Once an agent has write, send, delete, payment, deployment, export, or submission capabilities, its actions may directly trigger property loss, data leakage, contractual liability, labor disputes, consumer rights disputes, science and technology ethics risks, and regulatory liability.

For this very reason, when enterprises deploy AI agents, they cannot merely formulate a set of "AI Usage Rules," nor can they merely state in the system prompt that "high-risk actions must be confirmed by a human." When designing AI deployment, data compliance, and intelligent business process governance solutions for enterprises, the author has repeatedly encountered a very specific problem: enterprises do not lack the principle of "human in the loop"; what they truly lack is the design capability for "approval nodes." An approval node refers to a control step in which a qualified person reviews, confirms, rejects, modifies, escalates, or suspends before an agent executes a high-risk action, during execution, or before the result takes effect.

This article intends to discuss precisely how enterprises should set approval nodes around high-risk actions. The basic judgment of this article is: the key to AI agent governance is not to have enterprises approve all AI output, but to embed approval nodes into the agent's action chain according to action risk, data sensitivity, external impact, reversibility, ethical risk, and legal consequences, and to ensure through system permissions, tool invocation control, log traces, ethics review, and continuous auditing that they cannot be bypassed, can be traced, and can be held accountable.

01. Why can't we rely only on prompts to constrain agents?

In practice, many enterprises first adopt a low-cost approach: stating in system prompts or internal rules that "sensitive data must not be accessed," "formal documents must not be sent automatically," and "high-risk actions must be reported first." This approach is necessary, but it is insufficient to perform the compliance control function for high-risk actions.

First, prompts may be "forgotten." In long contexts, multi-turn tasks, multi-tool invocations, or multi-agent collaboration, an agent may gradually deviate from its initial system instructions and fail to continuously comply with the originally set approval requirements. Second, prompts may be "confused by noise." During operation, an agent simultaneously receives user instructions, system instructions, tool return results, external web page content, historical memory, and third-party data, and approval requirements may be diluted, overridden, or misinterpreted by other context. Even worse, prompts may also be maliciously injected. External web pages, emails, documents, knowledge bases, or third-party tool return content may embed instructions that induce the agent to bypass approval, leak data, invoke the wrong tools, or execute unauthorized actions.[1]

More importantly, prompts themselves lack enforcement power. Even if a prompt states "must not execute," if the corresponding tools and permissions remain open at the underlying system level, the agent may still complete the action under misunderstanding, malicious inducement, or abnormal conditions. The "Careful Adoption of Agentic AI Services" jointly published by the Five Eyes cybersecurity agencies points out that autonomous agentic AI inherits LLM-related risks, that external data sources may influence the agent's context through indirect prompt injection, and that organizations should not grant agents broad or unrestricted access permissions, especially not allow them to freely access sensitive data or critical systems.[2] Singapore's IMDA "Model AI Governance Framework for Agentic AI" also emphasizes that enterprises should limit the agent's action space through mechanisms such as access control, guardrails, human approval, logging, and monitoring, and cannot rely solely on prompt-layer control.[3]

Therefore, the necessity of approval nodes lies in this: they are not moral exhortations at the prompt level, but rigid controls at the system and management levels. High-risk actions must be systematically intercepted, submitted for approval, and recorded through permission, tool, process, and log mechanisms, and must be denied by default when approval is not obtained.

02. The relationship between approval nodes, human in the loop, and ethics review

"Human in the loop" is a common expression in AI governance, but when enterprises actually implement it, they need to break this principle down into specific approval nodes. "Human in the loop" answers whether humans should participate in oversight; "approval nodes" answer where humans participate, what they approve, on the basis of what information they approve, how the approval result is executed by the system, and how approval records are preserved and audited. "Human in the loop" without approval node design easily remains at the level of slogans; approval nodes without system enforcement easily degenerate into formalistic confirmation.

In addition, on March 20, 2026, the Ministry of Industry and Information Technology and nine other departments issued the "Measures for Ethics Review and Services of Artificial Intelligence Science and Technology (Trial)," incorporating the scope of application, implementing entities, working procedures, and supervision and management of AI science and technology ethics review into an institutional framework; relevant interpretations point out that ethics review focuses on aspects such as human well-being, fairness and justice, controllability and trustworthiness, transparency and explainability, traceability of responsibility, and privacy protection.[4] After the issuance of these Measures, when enterprises deploy AI agents, especially in scenarios involving human dignity, public order, life and health, labor and employment, consumer rights, fairness and justice, privacy and security, protection of vulnerable groups, or social and public interests, they cannot review only from the perspective of business efficiency and information security, but also need to incorporate the perspective of science and technology ethics governance.

The author suggests distinguishing the two as follows: ethics review addresses whether a certain type of AI agent scenario can go online, whether its scope can be expanded, and whether it can be granted greater autonomy; approval nodes address whether, when an agent is about to execute a certain high-risk action in a specific task, human approval is required, who approves, on the basis of what information approval is given, and how execution proceeds after approval. Ethics review is a scenario-level, pre-launch, and re-review governance mechanism; approval nodes are action-level, in-operation, and process-based control mechanisms.

The two should form a closed loop. Ethics review conclusions should be translated into approval nodes, permission restrictions, logging requirements, and continuous monitoring indicators; and approval records, rejection records, anomaly records, user complaints, and human intervention records should in turn become important evidence for ethics re-review. In other words, ethics review determines the risk boundary, and approval nodes enforce the risk boundary.

03. How to identify high-risk actions, and who approves them

Enterprises should not set human approval for all agent actions. Excessive approval leads to reduced efficiency, alert fatigue, and mechanical approval, which instead weakens governance effectiveness. A more reasonable approach is to first identify the actions an agent may execute, and then stratify them according to risk level, data sensitivity, external impact, reversibility, and ethical risk.

The first category is low-risk auxiliary actions, such as public information retrieval, organizing non-sensitive meeting minutes, ordinary document classification, internal knowledge base retrieval, and draft generation. Such actions usually do not involve sensitive data, do not change system states, do not directly affect third-party rights, and errors are easy to correct. Enterprises may allow automatic execution, but should retain logs, set up post-hoc spot checks, and establish user feedback mechanisms. The approval entity is mainly the business user department; the compliance or information security department may conduct sample inspections, and usually no ethics review committee involvement is required.

The second category is medium-risk content generation or internal suggestion actions, such as drafting emails, initial contracts, customer reply suggestions, internal reports, and compliance analysis suggestions. The risk in such scenarios is that AI results may be adopted by employees, but before formal external sending or effectiveness, humans still have the opportunity to review. Approval nodes should be set "before AI output is formally used." Ordinary business texts may be confirmed by the handling personnel or business responsible person; contracts, commitments, legal opinions, regulatory materials, or major customer communications should involve legal, compliance, or management in confirmation.

The third category is sensitive data processing actions, such as batch export of customer information, retrieval of employee salary and performance data, invocation of sensitive personal information, access to trade secrets, transmission of internal data to external tools, and use of data for retrieval augmentation or long-term memory. Once such actions are out of control, they may trigger personal information protection, data security, trade secret protection, and cybersecurity liability. Approval nodes should be set before data invocation, before export, before external transmission, or before third-party tool invocation. Ordinary business data is approved by the business responsible person to whom the data belongs; personal information involves the personal information protection officer or compliance department; sensitive personal information, batch personal information, core customer data, trade secrets, or important business data should be jointly approved by the business responsible person, legal, compliance, and information security. Where large-scale personal information processing, sensitive groups, major personal rights and interests, or data reuse is involved, the matter should be submitted to the ethics review committee or ethics governance body for scenario-level review.

The fourth category is external sending, commitment, or transaction actions, such as automatically sending quotations, confirming orders, sending contract texts, submitting bid documents, making refund or compensation statements, issuing external announcements, and submitting materials to regulators. Such actions may constitute the enterprise's declaration of intent, affect contract formation, performance, or interpretation, and may form evidence. Approval nodes should be set before external sending or before the result takes effect. Ordinary customer replies are approved by the business responsible person; those involving contracts, quotations, compensation, breach, termination, or dispute handling are approved by legal; those involving regulatory submissions, public releases, major transactions, or major customers should be jointly approved by the business responsible person, legal, compliance, and management.

The fifth category is funds, finance, and asset disposal actions, such as payments, refunds, reimbursements, purchase orders, account write-offs, automatic invoicing, asset disposal, or financial transactions. Such actions directly affect enterprise property, the cost of recovering errors is relatively high, and they may also involve internal control, auditing, and fraud risks. The author suggests that, in principle, fund payments, refunds, purchase confirmations, and account write-offs should not be completed end-to-end automatically by AI agents. Agents may organize materials, calculate amounts, flag anomalies, and generate process suggestions, but dual approval or special approval should be set before execution. Usually, the business handling department conducts preliminary review, the finance department conducts re-review, and amounts exceeding the threshold are approved by management; where disputes, compensation, or abnormal transactions are involved, legal should participate. Where high-impact scenarios such as financial consumers, personal credit, or user asset rights are involved, ethics review should also be considered.

The sixth category is system permissions, production environments and security configuration actions, such as modifying production code, executing scripts, modifying databases, deleting or overwriting files, elevating account privileges, modifying access control policies, deleting logs, and connecting new MCP Servers, plugins, external tools or APIs. Such actions may cause system interruption, data corruption, privilege proliferation or security vulnerabilities. Approval nodes should be set before tool invocation, before permission changes, before execution in the production environment and before log deletion. The IT or system head approves technical feasibility, the information security head approves security impact, and where business continuity is involved, the business head confirms; major system changes should go to the change committee or management for approval. Actions such as deleting logs, modifying audit records and self-elevating privileges should in principle be prohibited from autonomous execution by the agent.

The seventh category is actions affecting major personal rights and interests, such as recruitment screening, candidate ranking, performance evaluation, job adjustment recommendations, customer credit scoring, claim rejection, complaint handling, restriction of rights, and agent services involving minors, the elderly, persons with disabilities, psychological companionship, emotion recognition, anthropomorphic interaction or user dependence risk. Such scenarios readily involve transparency of automated decision-making, fair discrimination, human dignity and protection of vulnerable groups. The agent should be positioned as an auxiliary decision-making tool and should not directly make the final decision. Approval nodes should be set before AI recommendations are converted into final decisions, and channels for explanation, appeal, review and human correction should be retained. In HR scenarios, the HR head and the employing department confirm, and major labour matters are jointly reviewed by HR and legal; in consumer rights scenarios, the business head, customer service head, legal or compliance participate; where vulnerable groups, major personal rights and interests, fair discrimination or psychological impact are involved, ethical review or periodic re-review should be incorporated.

The eighth category is prohibited autonomous actions, such as deleting or modifying audit logs, bypassing security policies, self-elevating privileges, self-modifying permission boundaries, creating high-privilege accounts, connecting unknown third-party tools without approval, disclosing sensitive data without approval, and expanding the scope of use of high-risk agents without approval and without ethical review. Such actions should not be designed as "permitted for agent execution after approval", but should be prohibited at the system level from autonomous execution by the agent. Where exceptional handling is truly necessary, it should be jointly confirmed by the business head, legal, compliance, information security and management; where high ethical risk is involved, it should also be submitted for ethical review.

04. Mapping between approval subjects and scenarios

To avoid approval subject design becoming abstract, enterprises can establish an approval matrix of "scenario—action—node—person". This matrix is both a compliance design tool before agent deployment and a foundational document for subsequent system development, permission configuration, internal audit and external dispute evidence.

This matrix table transforms "who approves" from a vague management requirement into an executable organizational division of labour. When designing AI governance processes for enterprises, the author usually advises enterprises not to simply adopt the traditional approach of "approval by the direct supervisor", but to match approval subjects according to the nature of the risk: business risk is judged by the business head, legal consequences by legal, data processing by the personal information protection or data compliance head, security configuration by the information security head, financial risk by finance, and major ethical risk by the ethics review committee or ethics governance body.

05. How to specifically design approval nodes

The first step is to map the agent's action chain. Before designing approval nodes, enterprises should first sort out who triggers the agent, what input the agent receives, which data it accesses, which tools it invokes, whether it writes or deletes data, whether it sends information externally, whether it triggers transactions or system changes, whether results take effect automatically, whether rollback is possible after errors, and whether major personal rights and interests, public interest, fairness and justice, privacy and security or other ethical risks are involved. Only by mapping the action chain clearly can one determine whether approval nodes should be placed before task initiation, before data invocation, before tool invocation, before results take effect, when anomalies occur, or at the post-execution audit stage.

The second step is to establish an approval matrix of "action—risk—node—person—ethical review". The matrix should at least include business scenario, agent action, data involved, tools invoked, risk level, whether ethical risk is involved, whether ethical review is required, location of approval node, approval subject, approval materials, system action after approval, and logging and audit requirements. When enterprises carry out AI compliance projects in the future, this matrix can often be directly converted into lawyer deliverables, and can also serve as the implementation basis for technical teams to configure permissions and workflows.

The third step is to enable approvers to "understand the risk". An approval request should not merely pop up "agree or not". Approvers should at least see the agent's identity, triggering user, business scenario, proposed action, tools or systems invoked, types of data involved, whether personal information or trade secrets are involved, whether major personal rights and interests or vulnerable groups are involved, whether information is sent externally or third parties are affected, whether revocation or rollback is possible, the execution rationale given by the agent, key basis or sources, risk level, ethical risk warnings, and optional operations such as approve, reject, modify, escalate, suspend or submit for ethical review. Approvers of major risk actions should also fill in a brief rationale; otherwise, although approval records exist, it is difficult to prove that the approval was substantive.

The fourth step is to ensure that approval nodes are enforced by the system. Approval nodes should not exist only in institutional texts. The system should ensure that execution is not permitted without approval, that approval timeout defaults to rejection, that after rejection it is not possible to bypass by re-prompting, that approval is valid only for the current action, that permanent whitelists must not be set for high-risk actions, that re-approval is required after the action content changes, that re-approval is required for privilege elevation, that approval records must not be deleted or modified by the agent, and that approval content should be bound to the actual executed action to prevent "approving A, executing B". [5] Ethical review requirements must also not be replaced or bypassed by ordinary business approval.

This point is particularly critical for enterprises. Many compliance systems fail not because enterprises have not written rules, but because the rules have not been translated into system controls. This is especially true for approval nodes of AI agents: if approval is only written in rules, prompts or training materials, while the underlying tool permissions remain open, high-risk actions may still occur under abnormal conditions. A truly effective approval node should be a system-level control, not self-restraint at the prompt level.

From the perspective of legal service transformation, the design of AI agent approval nodes should not remain at verbal advice, but should form a deliverable, implementable and auditable document system. The author suggests that enterprises should at least form the following institutional documents and project deliverables.

Among the above documents, the "AI Agent High-Risk Action Approval Matrix" and the "AI Agent Action Risk Classification List" are the easiest for enterprise management to understand and the easiest to convert into business processes and system configurations; the "AI Agent Permission and Tool Invocation Whitelist" is the key bridge for converting legal requirements into technical permissions; the "AI Agent Ethical Review Guidelines" can help enterprises identify, under the new regulatory framework, which scenarios need to be submitted for ethical review, which scenarios only need preliminary compliance screening, and which scenarios must be suspended or have deployment restricted.

For lawyers, the value of such projects lies not only in issuing a legal opinion, but in helping enterprises complete the closed-loop design of "rules—processes—systems—evidence". In other words, lawyers should not only be risk warners, but should also become designers of AI governance structures.

07. Node management in deployment, operation and audit

Approval nodes are not a one-time design. Before, during and after the launch of AI agents, continuous management is required.

Before deployment, enterprises should complete use case risk assessment, data impact assessment, preliminary ethical risk screening, action list mapping, high-risk action identification, approval node matrix design, permission configuration review, logging scheme confirmation, system control testing, and rollback and emergency plan confirmation. Where necessary, it should be submitted to the ethics review committee or ethics governance body for review. The goal of pre-deployment review is to confirm that the agent will not come into contact with high-risk tools, sensitive data, critical systems or scenarios involving major personal rights and interests without the protection of approval nodes and ethical review.

During deployment, enterprises should not open all capabilities at once. A more prudent approach is to first conduct a small-scale internal pilot, starting from low-risk scenarios, limiting the scope of users, data and tools, and then gradually opening permissions. The focus of grey release is not only to test whether the agent can complete tasks, but also to test whether it will be correctly intercepted before high-risk actions, whether it will trigger approval, whether it will stop execution when not approved, and whether ethical review conclusions are implemented in system controls.

During operation, enterprises should focus on monitoring the number of approval node triggers, approval pass rate, rejection rate, approval response time, agent attempts to bypass approval, abnormal tool invocations, unauthorized access, bulk export, external transmission, repeated failures or loop invocations, as well as complaints, appeals, unfair results or abnormal outputs involving ethical risk. Upon discovering anomalies, automatic suspension, human review, permission downgrade, security investigation, node re-assessment, ethical re-review and rectification closed loop should be triggered.

Approval nodes themselves also need auditing. A long-term approval pass rate close to 100%, extremely short approval times, long-blank approval opinions, and the same approver concentrating on handling large numbers of high-risk actions may all indicate that approval has become formalistic. Audit indicators may include high-risk action identification accuracy, approval node trigger rate, approval rejection rate, number of unapproved execution events, post-approval error rate, log completeness rate, rectification completion rate, appeal rate in scenarios involving major personal rights and interests, proportion of agent recommendations overturned by humans, and implementation rate of ethical review conclusions. Audit and ethical re-review results should in turn be used to adjust risk classification, add or remove approval nodes, optimize the approval interface, adjust approval subjects, modify permission configurations, add system interception, conduct special training or suspend high-risk scenarios.

When an AI agent errs or causes damage, it is difficult for an enterprise to use "this is autonomous AI behaviour" alone as a ground for exemption. The truly meaningful defence is to prove that the enterprise has established reasonable risk identification, approval nodes, permission controls, ethical review, monitoring and auditing, and emergency correction mechanisms before deployment, during operation and after the event.

In personal information protection scenarios, approval nodes can help enterprises prove that they have identified the processing purpose and necessity, limited the agent's access scope, set additional controls for sensitive personal information, bulk export, external transmission and third-party invocation, and retained processing records and anomaly handling records. In data security and trade secret protection scenarios, approval nodes can prove that the enterprise has carried out data classification and grading, restricted the agent's access to core data, approved external transmission and tool invocation, and established access logs and anomaly alerts.

In contract and commercial liability scenarios, approval nodes can prove that AI-generated content does not automatically constitute the enterprise's final expression of intent; external quotations, commitments and contract confirmations must be approved by authorized personnel; agents may not confirm transactions beyond their authority; the enterprise's internal authorization chain is clear; and erroneous transactions have correction and accountability mechanisms. In human resources, consumer rights and technology ethics governance scenarios, approval nodes and ethics review can prove that AI agents serve only as auxiliary decision-making tools, that major rights-and-interests matters are ultimately decided by humans, that the enterprise retains review, appeal and correction mechanisms, and that it has assessed and re-reviewed automation bias, unfair outcomes, human dignity, protection of vulnerable groups and privacy security.

From the perspective of compliance evidence, the value of approval nodes lies not only in "preventing errors from occurring" but also in helping the enterprise prove that "even if an error occurs, the enterprise has fulfilled its duty of reasonable care." This is precisely what distinguishes approval nodes from ordinary process approvals: they are not a formal confirmation by management, but the evidentiary structure of the enterprise's AI governance responsibility.

Conclusion: Approval nodes are the institutional interface through which AI agents move from high-risk tools to trusted productivity

What AI agents bring is not merely an improvement in tool efficiency, but a redesign of corporate governance structure. Approval nodes are the key institutional interface connecting an agent's capacity for autonomous action with the enterprise's organizational responsibility; ethics review is an important governance mechanism for determining whether a high-risk agent scenario can be deployed, how it should be deployed, and whether deployment should continue.

The problem enterprises truly need to solve is not simply requiring that "humans participate," but clarifying: where humans participate, on the basis of what information they participate, with what authority they participate, which scenarios require ethics review, how approval results are executed by the system, and how approval and ethics review records are audited and used for accountability.

The author suggests that, before deploying AI agents, enterprises should treat approval nodes as a core governance object and design them in sync with permission management, data classification and grading, tool invocation control, ethics review, log trails, continuous monitoring and responsibility allocation. Only in this way can AI agents be transformed from a high-risk automation tool into a controllable, auditable and accountable enterprise productivity system that also meets the requirements of responsible innovation.

References

Questions and answers

Why are prompts alone insufficient to control high-risk actions by AI agents?

Prompts may be forgotten in long contexts, confused by noise, or maliciously injected through external content. They also lack enforcement power: if underlying tools and permissions remain open, the agent may still complete the action. Approval nodes must therefore be rigid controls at the system and management levels, denying execution by default without approval.

What is the difference between approval nodes and ethics review for AI agents?

Ethics review is a scenario-level, pre-launch and re-review governance mechanism addressing whether a type of AI agent scenario can go online and whether greater autonomy can be granted. Approval nodes are action-level, in-operation controls addressing whether a specific high-risk action requires human approval, who approves, and how execution proceeds. The two should form a closed loop.

Which AI agent actions should be prohibited from autonomous execution?

Prohibited autonomous actions include deleting or modifying audit logs, bypassing security policies, self-elevating privileges, self-modifying permission boundaries, creating high-privilege accounts, connecting unknown third-party tools without approval, disclosing sensitive data without approval, and expanding the scope of high-risk agents without approval and ethical review. These should be prohibited at the system level.

How should enterprises design approval nodes for AI agent actions?

Enterprises should first map the agent's action chain, then establish an approval matrix of action, risk, node, person and ethical review. Approvers must be enabled to understand the risk through detailed approval requests. Approval nodes must be enforced by the system, with no execution without approval, timeout defaulting to rejection, and binding approval content to the actual executed action.

How do approval nodes affect enterprise legal liability for AI agent errors?

It is difficult for an enterprise to rely solely on autonomous AI behaviour as a ground for exemption. Approval nodes help prove reasonable risk identification, permission controls, ethical review, monitoring and emergency correction. In personal information, data security, contract and ethics scenarios, they serve as compliance evidence that the enterprise fulfilled its duty of reasonable care.

Laws and cases cited

Regulations

  • Measures for Ethics Review and Services of Artificial Intelligence Science and Technology (Trial)

Policy documents

  • Careful Adoption of Agentic AI Services
  • Model AI Governance Framework for Agentic AI

References

  1. OWASP Agentic Security Initiative: "Agentic AI - Threats and Mitigations", https://genai.owasp.org/resource/agentic-ai-threats-and-mitigations/; "OWASP Top 10 for Agentic Applications 2026", December 9, 2025, https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/.
  2. Australian Cyber Security Centre (ASD ACSC) under the Australian Signals Directorate, U.S. Cybersecurity and Infrastructure Security Agency (CISA), U.S. National Security Agency (NSA), Canadian Centre for Cyber Security, New Zealand National Cyber Security Centre, UK National Cyber Security Centre: "Careful Adoption of Agentic AI Services", April 30, 2026, https://www.cisa.gov/resources-tools/resources/careful-adoption-agentic-ai-services; PDF address: https://media.defense.gov/2026/Apr/30/2003922823/-1/-1/0/CAREFUL%20ADOPTION%20OF%20AGENTIC%20AI%20SERVICES_FINAL.PDF.
  3. Infocomm Media Development Authority of Singapore (IMDA): "Model AI Governance Framework for Agentic AI, Version 1.5", May 20, 2026, https://www.imda.gov.sg/-/media/imda/files/about/emerging-tech-and-research/artificial-intelligence/mgf-for-agentic-ai.pdf.
  4. Ministry of Industry and Information Technology, National Development and Reform Commission, Ministry of Education, Ministry of Science and Technology, Ministry of Agriculture and Rural Affairs, National Health Commission, People's Bank of China, Cyberspace Administration of China, Chinese Academy of Sciences, China Association for Science and Technology: "Measures for Ethics Review and Services of Artificial Intelligence Science and Technology (Trial)" (MIIT Joint Science [2026] No. 75), issued and effective March 20, 2026; reposted page of the Ministry of Commerce policy database: https://policy.mofcom.gov.cn/claw/clawContent.shtml?id=105505; People's Daily report "Ten Departments Issue Measures for Ethics Review and Services of Artificial Intelligence Science and Technology", April 5, 2026, https://paper.people.com.cn/rmrb/pc/content/202604/05/content_30149307.html.
  5. Cyber Security Agency of Singapore (CSA): "Addendum on Securing AI Systems", 2026, https://www.csa.gov.sg/resources/publications/addendum-on-securing-ai-systems/.