Europe & Americas: UK AI Sovereignty Advice, Security Funding; Proposed US AI Investment Curbs; SEC Exam Priorities

On 19 October 2026, the UK’s Council for Science and Technology (CST) published its advisory letter, “AI sovereignty: preserving UK agency in an AI-enabled world”, setting out four recommendations on UK AI capability development, model access and supply dependencies.

This is advice from an independent advisory body, rather than an implementation plan adopted by the government. It recommends mapping the ecosystem and dependencies across compute, data, skills and models, designating a single accountable owner, and strengthening organisational, skills and data foundations. Its approach to models emphasises maintaining trusted access to leading global models while building selected domestic capabilities. The recommendations primarily address government and create no new enterprise procurement or data-localisation obligations.

UK public-sector suppliers and enterprises relying on overseas models should consider extending supply-dependency reviews to business continuity. In addition to recording model and cloud providers, identify who controls APIs, authentication, retrieval repositories, tool interfaces and usage licences. Test whether necessary data can be exported, configurations migrated and providers switched if a model becomes unavailable, a version is retired or a contract ends. After switching, revalidate business performance, permissions and sensitive-data flows, and define change notifications and exit support in contracts. These preparations help establish whether an enterprise has practical choices, beyond assessing deployment autonomy solely through server location or supplier branding.

The letter is dated 30 July 2026; this week’s development is its public release. It recommends public-interest pilots in areas such as national security, healthcare and education, comparing open-weight models, smaller domestic models and different training and access approaches, and considering allocations of national compute capacity for trials. Subsequent review should establish whether the government adopts the advice and how model selection, data access and responsibilities enter specific pilot arrangements. The publication does not establish a uniform UK enterprise deployment standard for “sovereign AI”.

On 24 October 2026, HM Treasury announced AI investment plans, including additional advanced-model security research funding and work to scope a regulatory sandbox supporting agentic cybersecurity deployments across critical infrastructure, the public sector and industry.

These are investment and policy arrangements. The government announced an additional £120 million for the AI Security Institute (AISI) for model testing, research and technical standards, alongside £60 million for the National Cyber Shield, including a £25 million challenge fund and work to scope a new regulatory sandbox. The sandbox remains at the design stage. The announcement does not specify participation conditions, testing boundaries or operating rules, and funding commitments do not establish that projects have launched or enterprises have deployment authorisation.

Teams providing cyber defence or critical infrastructure services, or planning to participate in government pilots, can prepare evidence of authorisation and recovery capabilities for agentic defence operations. For example, distinguish permissions for vulnerability scanning, remediation recommendations and direct system changes. Set approvals or hard execution boundaries for high-impact actions such as isolating business systems or changing production configurations, and rehearse credential revocation, task interruption and configuration rollback. Testing should assess false positives, incorrect remediation and service-disruption risks, alongside the number of vulnerabilities discovered. Procurement contracts can allocate evidence preservation, emergency response and recovery responsibilities among suppliers, operators and business teams. These are readiness recommendations for controlled testing, rather than uniform admission conditions imposed by this announcement.

The plans follow the UK’s AI risk defence measures announced on 12 October 2026, which included extending frontier AI cyber defence pilots and exploring identification, reporting and response arrangements for serious agentic incidents and near misses. The new announcement adds funding commitments and a proposed sandbox to support deployment, but provides no complete implementation timetable. Subsequent monitoring should focus on sandbox participation conditions, operating boundaries and evaluation arrangements. Pilot testing, production deployment and existing incident-reporting responsibilities should be managed separately, without treating pilot participation as satisfaction of all applicable requirements.

On 22 October 2026, the U.S. Department of the Treasury issued the proposed “Outbound Investment Security Program Regulations”, setting out changes to restrictions and notification arrangements for US persons’ investments involving covered foreign entities in AI and other technology sectors.

The proposal is not yet in force and would implement the Comprehensive Outbound Investment National Security Act (COINS Act). It would remove the existing notification category for AI transactions and broaden prohibitions on certain US investments in covered foreign entities developing AI. It also proposes controls in high-performance computing, hypersonics and biotechnology, alongside changes to covered persons, transactions and exceptions. Relevant US investors would determine whether a transaction is prohibited, notifiable or outside scope; the proposed mechanism does not establish transaction-by-transaction prior approval.

US investors and groups transacting through controlled foreign entities should consider including target companies’ actual research and development activities, ultimate controllers, affiliated entities and investment rights in due diligence before committing capital, establishing joint ventures or making further capital contributions. Transaction-risk assessments should avoid relying solely on incorporation location or previous training-compute thresholds. Where financing and technical cooperation proceed together, distinguish equity, convertible debt, joint ventures and procurement arrangements, and include provisions for information updates, reassessment following regulatory changes and exits where necessary. Ordinary model procurement does not automatically become a covered investment merely because it involves AI, but partnerships carrying equity or other investment rights require separate assessment. Coverage remains subject to transaction definitions and exceptions.

The existing outbound investment rules took effect on 2 January 2025 and continue to apply despite publication of the proposal. In addition to expanding AI restrictions, the proposal would bring Cuba, Iran, North Korea and Russia within the countries-of-concern framework, and introduce new transaction exceptions and a non-binding feedback process. Enterprises therefore need to update technology classifications, affiliated-entity screening and transaction assessments together. The public comment deadline is 7 December 2026; the final rules and effective date remain undetermined. Subsequent review should focus on the scope of AI development, boundaries for integrating third-party models, exceptions and the transition between the existing and new rules.

On 23 October 2026, the U.S. Securities and Exchange Commission (SEC) published its Division of Examinations’ “Examination Priorities: Fiscal Year 2027”, covering intelligent technology use, cybersecurity and automated investment services at regulated firms, including registered investment advisers and broker-dealers.

The document sets out the Division staff’s annual examination priorities; it changes no existing law and creates no new obligations. Examinations address oversight of intelligent technologies in fraud prevention, back-office operations, anti-money laundering, trading and best interest obligations, alongside cyber risks such as generated phishing and synthetic audio and video. For automated investment advice, priorities include whether representations are fair and accurate, operations and controls are consistent with disclosures, and algorithmic recommendations align with investor profiles or stated strategies. The relevant section uses the term “super intelligence” (SI); applicability still depends on the specific business and existing rules.

Firms providing automated investment advice or integrating AI into investment analysis, customer management and trading should consider mapping product claims and client disclosures to actual system configurations. Claims about human review, risk limits or continuous monitoring should be supported by corresponding approval records, configurations and operational evidence. After model or strategy updates, use representative client profiles to validate whether recommendations remain consistent with established investment objectives, and link records of versions, input data, key outputs and exception handling. For account changes or trading instructions potentially triggered by synthetic voice or video, test verification through an independent channel and escalation arrangements. These are examination-readiness recommendations; the report prescribes no uniform technical solution or log-retention period.

This examination focus has continuity: the SEC’s FY2026 examination priorities already addressed the accuracy of AI representations, technology oversight and consistency between automated investment advice and client profiles. FY2027 continues to cover these matters and explicitly identifies threats such as generated phishing and social engineering using synthetic audio and video, as well as technology oversight in areas including anti-money laundering and best interest obligations. The annual priorities inform self-assessment, while specific examinations remain based on firms’ activities and risks. Publicly available examination feedback and enforcement documents can subsequently help identify which disclosure and operational-control gaps have attracted specific regulatory attention.