Europe & North America: UK Studies Near-Miss Reporting; EU Seeks AI Safety Information and Cites a Quality Management Standard; New York Warns of Incorrect Chatbot Answers; FSB Finalises 12 AI Practices

On 12 October 2026, the UK Cabinet Office, together with the AI Security Institute (AISI), the National Cyber Security Centre (NCSC) and other agencies, announced measures to strengthen defences against AI risks, covering government cyber protection, advanced model evaluations and incident response for autonomous AI agents.

The announcement sets out policy measures: the government will expand frontier AI cyber defence pilots and confirmed that advanced model evaluations are underway. It is also developing plans to respond to serious incidents involving AI agents and exploring stronger protections through the Cyber Security and Resilience (Network and Information Systems) Bill and the wider regime, including identification and reporting of near misses, meaning events that narrowly avoid causing harm. The mechanisms remain under development. The Bill is still before Parliament, and the announcement introduces no uniform reporting deadline, threshold or channel.

Hospitals, utilities and their AI service providers can prepare by bringing unauthorised agent actions, abnormal sequences of tool calls and events that narrowly avoid service disruption into internal incident handling processes. We recommend recording authorisations, model versions, tool calls, the scope of potential impact and reasons for blocking high-impact actions intercepted by permission checks or stopped by humans, to help distinguish model errors, prompt injection and permission configuration issues. Joint deployment acceptance exercises with suppliers can test credential revocation, prevention of subsequent actions and transfer to human operators or existing processes. Contracts can allocate responsibility for evidence preservation, emergency contacts and support for recovery. These are preparation recommendations based on the policy direction and technical incident reviews; external reporting recipients and deadlines still need to be checked against currently applicable rules.

The Bill was introduced on 12 November 2025 to reform the existing network and information systems security regime. The new announcement identifies serious agentic AI incidents, loss of control and near-miss reporting as areas for further examination. Subsequent developments to verify include specific parliamentary amendments and how the government's response plans define triggers and allocate responsibilities for coordination. In the same week, the Cyber Security Agency of Singapore (CSA) published findings from an agentic AI security tabletop exercise on 15 October 2026, recommending runtime monitoring, hardened execution environments and access controls, without introducing mandatory obligations. The UK is examining government response capabilities alongside institutional design. Organisations operating across jurisdictions can reuse exercises and evidence records while maintaining separate local rules for incident escalation and external reporting.

On 12 October 2026, the European Commission published a follow-up announcement on the special meeting of the AI Scientific Panel, confirming that the European AI Office (AI Office) had requested and received serious AI incident reports and had sent requests for information on safety practices to the relevant model developers.

The announcement describes progress in regulatory information gathering and expert consultation; it does not announce findings of infringement or penalties. The Commission will combine questions from the AI Scientific Panel and the AI Office for further enquiries. Experts also recommended governance of internal models during development and testing, stronger incident investigation, and scrutiny of training techniques and architectures that reduce model monitorability, meaning the ability to observe behaviour and identify anomalies. The expert recommendations do not yet create new uniform obligations. The announcement does not publish company names, the full requests or specific response deadlines.

Enterprises supplying models to the EU or building network-connected agents using external models should consider preparing incident evidence that can be independently reviewed. During development and testing, define the scope of network connectivity, code execution and tool access, and link records of model versions, task objectives and permission changes. When anomalies occur, preserve call traces, alert timestamps, containment measures and retest results. Buyers can agree that suppliers will provide evaluation summaries, incident notifications and investigation support, and specify which operational evidence each party will preserve. Where monitoring evidence is insufficient, enterprises can restrict autonomous execution and undertake additional validation before deciding whether to expand deployment. These are deployment and procurement recommendations; specific reporting responsibilities still depend on the organisation's role, model classification and applicable provisions.

The special meeting on 9 October 2026 focused on loss-of-control risks. The follow-up announcement adds information about incident reports and enquiries already undertaken. Further developments to verify include whether investigations lead to specific evaluations or corrective measures. Article 55 of the Artificial Intelligence Act (AI Act) establishes serious incident documentation and reporting obligations for providers of general-purpose AI models with systemic risk. Requests for information under Article 91 must specify their legal basis, purpose, the information sought and the response deadline; applicability still requires assessment of model classification and transitional arrangements. Measures announced by the UK Cabinet Office on the same day continue to examine incident and near-miss reporting for AI agents, while the EU announcement discloses information gathering already underway. Enterprises operating across jurisdictions can reuse incident evidence while separately maintaining reporting responsibilities, regulatory contacts and procedures for responding to requests.

On 14 October 2026, the European Commission published Implementing Decision (EU) 2026/2274 in the Official Journal of the European Union, publishing the reference to harmonised standard EN 18286:2026, “Artificial intelligence – Quality management system for EU AI Act regulatory purposes”, concerning quality management systems for providers of high-risk AI systems.

The Decision was adopted on 12 October, published on 14 October and entered into force on the same day. The central change is that conformity with the standard or relevant parts can confer a “presumption of conformity” with the AI Act requirements covered, helping providers demonstrate that the corresponding requirements have been met. The standard translates quality management requirements into implementable, auditable processes. However, the presumption has defined limits and does not establish that an entire AI system meets every AI Act requirement. Its precise scope still needs to be checked against the formal standard and its mapping to regulatory requirements.

For enterprises providing high-risk AI systems to the EU market, we recommend first confirming their regulatory role and system classification, then establishing a mapping between “AI Act requirement—standard clause—system version—person responsible for evidence”. Drawing on the quality management requirements in Article 17 of the AI Act, integrate data processing, validation and testing, model changes, release approvals, post-market monitoring and incident handling into development and production processes. Specify who operates each control, what records it produces and how it is reassessed after changes. Buyers should seek clarification from suppliers on the products and versions covered by their compliance statements and the requirements that still need separate evidence. They can also agree arrangements for change notifications and evidence delivery, rather than approving deployment solely on the basis of a management system statement. These are implementation recommendations based on the existing Act and the standard's citation; the Decision does not introduce a uniform logging period or approval procedure.

The standard had already been published in July 2026. The new development this week is its citation in the Official Journal and the corresponding presumption-of-conformity effect. For integration with existing sectoral arrangements, Article 17(3) allows providers already subject to EU sectoral quality management obligations to incorporate AI-related requirements into their existing systems. Enterprises can use this approach to assess gaps and add AI-specific controls. Implementation timing needs separate management: under the current Article 113, the relevant requirements apply to Annex III high-risk systems from 2 December 2027, and to product-related high-risk systems covered by Annex I from 2 August 2028. Enterprises should continue to check new harmonised standard references and their coverage, maintaining separate plans for standards adoption and the statutory application of obligations.

On 14 October 2026, the Office of the New York Attorney General (OAG) released an updated guide, “Protecting New York Voters from AI-Generated Election Misinformation”, advising voters on identifying deepfakes, verifying voting information and reporting concerns.

This is an update to public risk guidance. The announcement introduces no new enterprise technical specification or enforcement decision. The guide emphasises verifying voting information against official sources: chatbots may provide information for the wrong location, incorrect deadlines or false voting requirements, while deepfakes may impersonate government notices or officials. Visual flaws and AI detection tools cannot be relied on to establish authenticity. OAG advises consulting state or local boards of elections and reporting suspicious content to platforms and the office. The announcement does not identify chatbot products, error rates or a testing methodology for this update.

Teams providing public question-answering, search summaries or election content services should consider incorporating location, timeliness and source verification into release controls. For poll sites, eligibility and deadlines, prioritise information from the relevant state and county election authorities, and link answers to the original page, applicable jurisdiction and verification time. When location is unclear, sources conflict or information cannot be verified, direct users to official lookup tools instead of allowing the model to fill in critical facts. Pre-deployment testing should cover places with the same name, differences between early voting and Election Day rules, outdated pages and multilingual questions. Revalidate after changes to the model or retrieval sources. During operation, maintain linked records of answers, cited sources, versions and corrections, so complaints can trigger review and content updates. These are deployment recommendations based on the risks described in the guide; the update does not make these particular configurations mandatory.

The warning has a testing basis: on 1 November 2024, OAG published sample testing of chatbots that identified incorrect voter registration deadlines and poll-site advice. The update carries those warnings into the 2026 election, with the next announced service milestone being the opening of the Election Protection Hotline on 24 October 2026. Alongside these public alerts (2024 testing announcement; 2026 printable guide), section 14-106(5) of New York's Election Law already establishes disclosure obligations for political communications knowingly containing materially deceptive media, with exceptions for satire, specified news reporting and a platform's initial dissemination of third-party content, among others. Enterprises therefore need to assess political communication disclosure responsibilities separately from accuracy controls for question-answering systems. An AI label does not resolve the problem of incorrect voting information.

On 15 October 2026, the Financial Stability Board (FSB) published “Sound Practices for Responsible Adoption of Artificial Intelligence (AI): Final report”, setting out 12 practices for all types of financial institutions, covering organisation-wide governance and the AI lifecycle.

The report is final but does not constitute a binding international standard. Institutions can tailor adoption to their size, complexity and use-case risks. The first four practices address strategic oversight, governance and accountability, risk management and documentation, and organisational adaptability. The remaining eight cover materiality assessment, model selection, data governance, explainability and transparency, performance management, human oversight, cyber and information and communication technology risks, and third-party risk. For agents, the report adds arrangements combining human oversight with deterministic controls, including code-based restrictions on tool access, triggers for human approval, and testing of safe stopping and permission revocation mechanisms.

Institutions using agents in payment, trading or lending workflows can prioritise checking whether authorisation, execution and recordkeeping are connected end to end. We recommend linking use-case owners, data and tool permissions, risk assessments and test results within deployment approval. For payments, trades or account changes, define prohibited actions, thresholds for human approval and conditions for stopping anomalous activity, then rehearse permission revocation, task handover and recovery. Procurement reviews should extend to embedded AI in software, subcontracted services and dependencies on critical suppliers, clarifying change notifications, investigation support and assistance with exit. Operational records can link authorisations, key tool calls, human interventions and execution outcomes according to risk, enabling anomalous actions to be reconstructed. These are deployment recommendations based on the toolkit; specific configurations still depend on business risks and local rules.

The work began with the consultation report on 10 June 2026. The final report incorporates around 160 responses, further clarifies the respective roles of boards and senior management, adds considerations on independent validation, and expands the treatment of embedded AI, subcontracting and supplier concentration risks. FSB also notes that the additional agentic AI considerations have not undergone a separate, full consultation. These practices remain under development, and FSB will continue monitoring and updating them as needed. At jurisdictional level, the Monetary Authority of Singapore (MAS) issued its “Guidelines on Artificial Intelligence Risk Management for Financial Institutions” on 7 October 2026, setting out local supervisory expectations and taking effect from 7 October 2027. Cross-border groups can use the FSB practices as a common control reference while separately maintaining applicable regulatory requirements and implementation schedules for each entity.