EU Refines GPAI Implementation, UK Opens Legal AI Sandbox, U.S. Advances Industrial Agents and Investment Restrictions, Korea and Singapore Focus on Agentic AI Governance

On August 3, 2026, the European Commission’s AI Office published information on the fourth meeting of the Signatory Taskforce for the General-Purpose AI (GPAI) Code of Practice, further explaining how providers of GPAI models with systemic risk should conduct post-market monitoring of real-world model use and how signatories should disclose information about web crawlers and copyright rights reservations. It is important to note that the meeting itself took place on July 17, 2026; August 3 was the date on which the Commission published its meeting summary. The activity formed part of implementation coordination for the Code of Practice rather than the adoption of new legislation or additional regulatory rules.

The discussion covered both the Safety and Security Chapter and the Copyright Chapter of the GPAI Code of Practice. On safety, the AI Office explained under Measure 3.5 that analysing how a model is actually used and behaves after being placed on the market can serve as supplementary evidence for post-market monitoring and systemic-risk assessment, complementing pre-deployment model evaluations in subsequent risk determinations. The meeting also discussed “marginal-risk clauses” in safety and security frameworks. The AI Office made clear that such clauses should be used only in exceptional circumstances and with appropriate evidence and procedural safeguards. On copyright, under Measure 1.3（4）, signatories commit to publishing information about the web crawlers they use, robots.txt functionality, and other measures taken to identify and comply with rights-reservation expressions, and to provide affected rightsholders with an automated means of receiving updates when that information changes. Two levels of legal effect need to be distinguished. These specific practices are commitments assumed by signatories under the Code of Practice. By contrast, Article 53 of the AI Act imposes a legal obligation on GPAI model providers to put in place a policy to comply with EU copyright law and to identify and comply with rights reservations expressed for text and data mining.

For providers of models with systemic risk that use the Code of Practice as a compliance pathway, model governance can no longer treat “pre-deployment testing completed” as the endpoint. The more practical deployment question is whether the enterprise can continuously identify how the model is being used, what anomalous behaviour emerges in real-world use, and how those signals relate to specific model versions, deployment channels and existing risk scenarios, so that systemic risk can be reassessed. Enterprises can accordingly design model-use classifications, anomalous-behaviour records, incident-escalation procedures and version-traceback mechanisms, but these technical arrangements remain implementation choices rather than uniform log fields newly mandated by this taskforce meeting. The copyright side has a more direct effect on training-data acquisition: signatories that use web scraping for training need to maintain publicly accessible information on their crawlers and their handling of rights reservations, and ensure that rightsholders can be notified of changes. Post-deployment risk evidence and copyright evidence relating to training-data acquisition are becoming two separate compliance chains, both of which require continuous traceability. GPAI providers that do not sign the Code are not exempt from the AI Act’s obligations; they must demonstrate compliance through other sufficient means. The Commission has also indicated that, because regulators will have less information about their control measures, such providers may face more requests for information or requests for access for model evaluation.

This taskforce discussion came immediately after a clearly defined enforcement milestone. The AI Act’s obligations for GPAI model providers have applied since August 2, 2025, while the European Commission’s enforcement powers in relation to those obligations began to apply on August 2, 2026. The GPAI Code of Practice is a voluntary compliance tool endorsed by the Commission and the AI Board, allowing signatories to demonstrate compliance with relevant obligations under Articles 53 and 55. The information published on August 3 is therefore best understood as continued clarification of concrete implementation methods after the rules entered the enforcement phase, rather than as an additional layer of standalone regulation. Compared with the Artificial Intelligence Technology Evaluation program launched by the U.S. National Institute of Standards and Technology in July, the two mechanisms address different stages of model evaluation. NIST AITE mainly uses blind test data and a sequestered environment to improve the credibility of pre-deployment performance evaluation, while the EU discussion explicitly brings real-world usage information into post-deployment systemic-risk assessment. For cross-regional model providers, these mechanisms can correspond respectively to pre-launch validation and post-launch monitoring, but voluntary U.S. evaluation cannot substitute for the legal obligations already borne by GPAI model providers in the EU.

On August 3, 2026, the UK Department for Business, Innovation, Science and Trade formally opened applications for the Legal Services Advisory AI Growth Lab, providing developers and deployers of legal AI products with a multi-regulator consultation and testing mechanism. Applications close on September 27, 2026.

The mechanism is an advisory regulatory sandbox and does not change existing law or provide regulatory exemptions. The Legal Services Board (LSB), Solicitors Regulation Authority (SRA), Council for Licensed Conveyancers (CLC) and Information Commissioner’s Office (ICO) will jointly address cross-regulatory questions involving client confidentiality, legal professional privilege, personal data, explainability, professional responsibility and the use of client data. Applicant projects must have AI at the core of the regulatory question and must already have reached a certain level of development maturity. The first cohort is expected to include approximately 10–12 participants, with selected projects able to participate for up to nine months. The Lab cannot amend the law, provide a safe harbour or replace formal regulatory approval; participating enterprises remain responsible for complying with existing legal and regulatory obligations.

In practice, deploying legal AI often engages professional rules and data-protection obligations simultaneously, which is the most direct governance value of the Lab. For example, when a law firm uses historical client files to train or fine-tune a model, it may need to assess whether the information can be used for a new processing purpose, whether confidentiality or legal professional privilege applies, and what data the model supplier may access. If a tool directly analyses conveyancing materials or provides legal services to clients, the enterprise also needs to determine who reviews incorrect outputs, which results require lawyer intervention and how system limitations are communicated to users. UK government materials have already identified “training on historical client data” and “automated analysis of property-sale materials” as representative cross-regulatory scenarios. Enterprises can therefore use the Lab to clarify data-use, human-oversight and responsibility boundaries before formally scaling deployment, while still maintaining their own records of data sources, model versions, output reviews and regulatory communications during testing. These are deployment preparations derived from the mechanism rather than new uniform legal obligations created by the Lab.

The UK government first announced the Advisory AI Growth Lab on June 8, 2026, with legal services selected as the first sector. The change on August 3 was that the mechanism moved from policy announcement into formal application and project selection. Its institutional focus is also more restrained than the original cross-sector concept of an AI Growth Lab: the current legal-services Lab primarily helps enterprises understand how existing rules apply together to new AI products, and explicitly does not provide legal exemptions. This creates a relatively clear regional contrast with the European Union. The EU AI Act establishes legal obligations for different risk categories of AI systems through horizontal legislation, whereas the UK is initially using coordination among sector regulators to address overlapping rules and uncertainty about how they apply. For cross-regional legal-technology companies, the same AI product may therefore face different governance tasks. In the UK, the immediate priority is to clarify how existing professional regulation and data-protection requirements intersect; in the EU, the enterprise must additionally determine whether the system falls within a specific AI Act risk category and what obligations follow.

On August 4, 2026, the U.S. National Institute of Standards and Technology (NIST) signed a memorandum of understanding with the U.S. Department of Energy Office of Science to join the Genesis Mission and apply AI agents to two initiatives involving advanced manufacturing and critical-infrastructure cybersecurity. This is a federal government collaboration initiative, not a regulatory rule, technical standard or mandatory enterprise obligation.

NIST will primarily advance the work through two previously established AI centers focused on economic security, carrying out two-year applied projects. The manufacturing track will develop and adopt AI-driven autonomous agents in combination with human-in-the-loop robots and autonomous systems—meaning that AI may independently execute certain tasks while retaining human participation or control at critical points. The first project uses civilian and military drone production as the initial scenario, with a goal of expanding production capacity to ten times the current level within two years. The critical-infrastructure track will develop AI agents for ultra-high-speed detection and remediation of cyber threats across power grids, telecommunications networks, water-treatment facilities, financial platforms and healthcare systems. Both initiatives aim to produce capabilities suitable for commercial and industrial deployment and to extract methods that can transfer to other manufacturing and cybersecurity scenarios.

For manufacturers and critical-infrastructure operators, the main difference between these agents and ordinary internal chatbots is that their outputs may directly translate into equipment control, production scheduling or network-response actions. Governance therefore shifts further toward “what the agent is allowed to do, and under what circumstances control must return to a human.” Enterprises drawing on these projects for autonomous-agent deployments should tier tool and system permissions: which actions may only read information and make recommendations, which may execute automatically, and which—such as shutdowns, configuration changes, network isolation or control of physical equipment—require human authorisation. Cybersecurity agents should record the basis for threat assessments, tools invoked, actions executed, permission identities and execution results, while also defining stop and rollback paths for false positives. Manufacturing deployments should connect model outputs with equipment safety interlocks, human operating authority and procedures for handling production anomalies. These are enterprise-governance preparations inferred from the projects’ application models rather than uniform control requirements imposed by NIST through the memorandum of understanding.

NIST’s participation in the Genesis Mission does not mark a sudden entry into industrial-agent research. In December 2025, NIST had already committed $20 million to work with the MITRE Corporation to establish two AI centers focused on manufacturing and critical infrastructure, including development, evaluation and adoption of AI agents for manufacturing operations and cybersecurity. In April 2026, NIST also launched work on an AI RMF Profile on Trustworthy AI in Critical Infrastructure, intended to develop more specific risk-management practices for AI agents and tools used in critical infrastructure. The change on August 4 is that these existing projects are now further connected to the Department of Energy-led Genesis Mission, which links national laboratories, supercomputing facilities, AI systems and scientific data for national-scale missions in energy, scientific discovery and national security. Unlike the European Union’s path of establishing legal obligations primarily through the AI Act, the U.S. approach here is closer to “develop and validate agents first in high-value, high-risk real-world environments, while building evaluation and risk-management tools in parallel.” Enterprises can draw on the control questions emerging from these projects, but should not treat Genesis Mission project requirements themselves as new U.S. AI compliance obligations.

On August 5, 2026, the Federal Register published a notice signed by the President on August 3 extending for one year, from August 9, 2026, the national emergency underlying Executive Order 14105 concerning the development of sensitive technologies in “countries of concern”. The notice preserves the legal basis for the U.S. Outbound Investment Security Program and does not add new countries of concern, technology categories or AI computing thresholds.

The rules that currently govern enterprise transactions remain the final regulations issued by the U.S. Department of the Treasury under Executive Order 14105, effective since January 2, 2025: Provisions Pertaining to U.S. Investments in Certain National Security Technologies and Products in Countries of Concern (31 CFR Part 850). The current regime identifies China, including Hong Kong and Macau, as a “country of concern”, and covers semiconductors and microelectronics, quantum information technologies and artificial intelligence. Covered transactions are divided into prohibited transactions and notifiable transactions. For AI, where a transaction does not otherwise fall within the prohibited category, an AI system trained using more than 10^23 computational operations, or a system intended for specified uses involving military applications, government intelligence, mass surveillance, cybersecurity, digital forensics, penetration testing or robotic control, may fall within the notification regime. Transactions involving AI systems specifically designed or intended for military, government intelligence or mass-surveillance uses, or general AI systems trained using more than 10^25 computational operations, as well as AI systems trained primarily on biological sequence data using more than 10^24 computational operations, may be prohibited. The extension of the national emergency does not alter these existing categories or thresholds.

The regime affects enterprises first at the investment-decision stage rather than after model deployment. Before investing in an AI company in mainland China, Hong Kong or Macau, U.S. investors cannot rely only on the target company’s industry label; they need to understand the actual AI systems under development, intended uses, training compute and relationships with covered persons. Covered transactions extend beyond direct equity investment and may include convertible interests, debt financing that provides governance rights analogous to equity, certain greenfield projects, joint ventures and interests in certain non-U.S. investment funds. U.S. persons also have obligations relating to controlled foreign entities, including preventing prohibited transactions and notifying Treasury of certain transactions. Investment committees and M&A teams therefore need to embed AI technical due diligence into transaction processes by requiring target companies to explain model use, training scale, primary data types and technical architecture, and by including representations and warranties, ongoing information rights and business-change notification clauses in transaction documents. For companies planning to continue scaling models, the fact that the relevant computing threshold has not been reached at the time of the initial investment does not mean that a later equity conversion or follow-on investment will necessarily remain in the same category. The rules assess different transaction points separately and require U.S. persons to evaluate the information they should reasonably possess under the “reasonable and diligent inquiry” standard.

This governance track began in August 2023, when Executive Order 14105 first declared the relevant national emergency and directed Treasury to establish an outbound investment restriction regime. Treasury issued the final rule in October 2024, and it became effective in January 2025; the August 2026 notice continues the national emergency that supports the program. Compared with the European Union’s current approach, both jurisdictions include artificial intelligence, advanced semiconductors and quantum technologies in outbound-investment economic-security discussions, but the regulatory intensity differs. The European Commission’s 2025 recommendation on outbound investment primarily called on Member States to review and assess technology-leakage risks and adopted a country-neutral review scope, while the United States has established directly applicable prohibited and notifiable transaction rules focused on China. Cross-regional corporate groups therefore cannot rely on a single “AI investment risk” label. They need to assess separately the nationality of the investor, the location of the target company, the target’s model-related technical activities and the transaction structure, and then determine whether the appropriate response is an internal risk assessment, a regulatory notification or non-participation in the transaction.

On August 5, 2026, Korea’s Personal Information Protection Commission (PIPC) announced the launch of a reform process for the personal information framework in the AI era. Beginning August 6, it invited policy proposals from the public, businesses, research institutions, academia and civil-society organisations on issues including AI agents, cross-border use of pseudonymized data and continuous collection of personal information by physical AI. The initiative is an early-stage public consultation preceding institutional reform, not an amendment to the Personal Information Protection Act, a regulatory draft or a new obligation already in force.

The reform is not aimed at any single AI product. Instead, it re-examines whether existing personal information rules can adapt to new forms of data processing. PIPC identified four concrete issues. First, personal information processing still relies heavily on repeated and formalistic individual consent. Second, international collaborative research using pseudonymized medical and other research data may face practical difficulties if fresh consent is required at the cross-border transfer stage. Third, when multiple AI agents collaborate or invoke external services, repeated new processing activities and consent questions may arise, creating new needs for allocating responsibility. Fourth, physical AI such as robots, drones and smart glasses continuously collects information from the surrounding environment, raising new questions for existing personal-information principles, safeguards and data-subject rights. PIPC established a “Personal Information System Reform Task Force” on July 30, with submissions open through August 31. It plans to narrow the reform agenda through public discussions and seminars and publish institutional-reform directions within the year. No specific legislative amendment has yet been produced.

For enterprises already deploying AI agents, the most useful preparation at this stage is to map “where the data goes as the task is executed”. An ordinary chatbot typically receives a single user-submitted input, whereas an agent may read email, calendars or customer records and then invoke search, payment or other agents to complete a task. Each additional tool call can introduce a new recipient, processing purpose and responsible party. Enterprises can therefore begin by building agent data-flow maps that connect user requests, accessed data, external models or tools, returned results and storage locations, while identifying which actions can rely on existing authorisation, which may require renewed user notice, and which party bears responsibility if data is mishandled or leaked. Cross-border research or model-development scenarios should separately record pseudonymization, overseas recipients, transfer grounds and re-identification restrictions. Physical AI such as robots and smart glasses should also distinguish between data actively submitted by users and images, audio or location data continuously collected by devices. These are deployment preparations responsive to potential reform directions, not new controls that PIPC has already required enterprises to implement immediately. Enterprises remain subject to the existing Personal Information Protection Act and current regulatory guidance when determining lawfulness.

The consultation continues a multi-year Korean policy trajectory on AI and privacy governance. In 2023, PIPC issued policy directions for personal information in the AI era, beginning to use principle-based guidance, regulatory sandboxes and “Prior Adequacy Review” to address legal uncertainty arising from new technologies. Its 2025 guidance on personal-data processing for generative AI already incorporated agent management across development and use lifecycles. The Third Master Plan for Personal Information Protection to Promote Trust-Based AI Innovation (2027–2029), announced in July 2026, further proposed studying accountability for agent decisions, continuous collection by physical AI and cross-border data-transfer mechanisms. The change in August is that the discussion is moving from how to interpret existing law toward whether the underlying consent, responsibility and cross-border transfer structures themselves need adjustment. Compared with the European Union, the institutional path also differs. The General Data Protection Regulation (GDPR) already provides legal bases beyond consent, including contractual necessity and legitimate interests, and uses adequacy decisions, standard contractual clauses and binding corporate rules for cross-border transfers. Korea is now reconsidering how its existing regime should adapt to AI agents, multi-party collaboration and pseudonymized international research. Cross-regional enterprises therefore cannot simply replicate a standard “user consent pop-up”; they need to assess separately the available processing grounds, cross-border mechanisms and agent-accountability structures in each jurisdiction.

On August 5, 2026, the Monetary Authority of Singapore (MAS) stated in a written parliamentary reply that its proposed Guidelines on Artificial Intelligence Risk Management apply to all uses of AI by financial institutions, including agentic AI. MAS also continues to follow a principles-based regulatory approach and has not announced a separate mandatory regulatory regime specifically for agents.

A distinction needs to be drawn between the proposed supervisory framework and technical implementation tools. MAS consulted publicly on the Guidelines on Artificial Intelligence Risk Management in November 2025, proposing supervisory expectations for financial institutions across board and senior-management oversight, risk-management frameworks and AI lifecycle controls. The scope included generative AI and AI agents from the outset, and the Guidelines have not yet been finalised. Separately, in July 2026 MAS published Safeguards for Agentic Finance at Runtime (SAFR), an industry-led effort exploring runtime controls including policy-constrained execution, real-time validation and auditability. The August 5 reply did not announce that SAFR would become a mandatory technical standard; instead, it placed SAFR within a set of practical implementation resources that sit alongside regulatory expectations.

For banks, insurers, payment institutions and capital-markets firms, the more immediate implication is that a system should not sit outside the existing AI risk-management framework merely because it is described as an “agent”. Ordinary generative AI may only produce recommendations, whereas an agent may go further by invoking payment, trading, customer-management or internal operational tools, meaning that model risk can translate directly into operational risk. Enterprises can accordingly include agents in a unified AI inventory and risk-tiering system while managing their action permissions separately: which tools are read-only, which actions may execute automatically, and at what thresholds high-impact operations such as transfers, trades or changes to customer entitlements require human authorisation. The audit trail needs to capture not only model outputs, but also which tools the agent invoked, under whose identity it acted, what actions it executed and when human takeover was triggered. Financial institutions using third-party models, external plugins or multi-agent collaboration also need to include the data-access rights and execution permissions of external components in supplier assessments. These are governance preparations derived from MAS’s published supervisory direction and the SAFR approach, not a new mandatory control checklist introduced on August 5.

The reply continues Singapore’s gradual progression from general AI governance frameworks toward runtime controls for financial services. In November 2025, MAS first proposed financial-sector risk-management guidelines spanning the entire AI lifecycle. In January 2026, the Infocomm Media Development Authority (IMDA) published the cross-sector Model AI Governance Framework for Agentic AI and updated it again in May. In July, MAS added SAFR specifically for financial agents. On August 5, it clarified that financial agents remain within the proposed general financial-sector AI risk framework rather than awaiting a standalone “Agent regulation”. Compared with the European Union, Singapore currently places greater emphasis on principles-based institutional oversight and runtime controls for financial institutions. The EU AI Act, by contrast, mainly determines high-risk status according to specific use cases, such as creditworthiness assessment and certain risk assessments in life and health insurance; being an “agent” is not itself a separate high-risk category. For the same financial agent deployed across regions, Singapore therefore places greater emphasis on the question “how does the institution control its autonomous behaviour?”, while the EU also requires an earlier determination of “which legal risk category applies to the specific financial use being performed?”