Europe & US: NYC Holds AI Safety Hearing; U.S. Lawmakers Reintroduce CISA Securing AI Task Force Bill; UK Accepts All 44 Healthcare AI Commission Recommendations; Canada Updates Call Recording and Voiceprint Consent Guidance; EU Convenes Special Meeting on Frontier AI Safety

On 5 October 2026, the New York City Council's Committee of the Whole held an oversight hearing on artificial intelligence risks, hearing testimony from AI companies, experts, and whistleblowers and considering legislative proposals on model validation, liability for harm, and deepfakes.

This was an oversight hearing and consideration of proposed legislation; the proposals have not been passed. Int. 1100-2026 would make third-party validation and human-initiated shutdown capabilities conditions for marketing, selling, and deploying models in the city, with validation covering performance, bias, data handling, and security. Int. 1116-2026 would allow victims to sue model providers that failed to implement reasonable safeguards against foreseeable harm from third-party misuse occurring in the city. Int. 0504-2026 would restrict the generation of deceptive audio, video, and images in response to notices withholding authorization from public officials and candidates, with a notification window within the 120 days before an election and a restriction on generation during the 60 days before an election.

Businesses providing models, offering generative services, or procuring related systems in New York City should prepare evaluation and control evidence tied to each deployment. Link supplier evaluations, business use-case tests, and human-initiated shutdown exercises to model versions, and define who can shut down a system, the scope of shutdown, and approval for resuming operation. When integrating external models, verify whether the business can block calls, revoke tool permissions, and retain response records. For services accepting unrestricted public input, test attempts to bypass safeguards, impersonation, and malicious generation; document identified risks, measures taken, and remaining limitations to support an assessment of whether safeguards are reasonable. Teams offering audio and video generation depicting individuals can also prepare processes for receiving notices, verifying identity, applying restrictions, and handling appeals. These are deployment preparation recommendations based on the proposals, rather than mandatory obligations introduced by the hearing.

In 2025, New York City passed laws governing municipal AI use, establishing algorithmic accountability review, standards for development and procurement, and a public inventory. This hearing extended the discussion to commercial model sales and deployment and liability for third-party misuse. Council records show that the validation and liability proposals were heard as preconsidered items on 5 October 2026 and formally introduced on 8 October 2026. They remain in committee; subsequent text revisions and votes should be tracked. Int. 1109-2026, considered at the same hearing, would use city procurement contracts to require contractors to notify the Office of Cyber Command (NYC3) within 24 hours of becoming aware of a reportable AI safety incident. These proposals address separate areas: conditions for commercial deployment, civil remedies, and public procurement terms. Businesses serving both commercial clients and city government should identify the products, contracts, and responsible entities within each proposal's scope, rather than combine the proposed controls into a single set of rules assumed to be in force.

On 6 October 2026, U.S. Representatives Troy A. Carter and Bennie G. Thompson reintroduced the CISA Securing AI Task Force Act (H.R. 10752), proposing an artificial intelligence task force within the Cybersecurity and Infrastructure Security Agency (CISA).

The bill has been referred to the House Committee on Homeland Security and has not become law. According to the sponsors, its central mechanism is internal coordination of AI safety and security activities within CISA. Within one year of enactment, a task force comprising personnel from different agency offices and divisions would coordinate the safe and secure design, development, adoption, and deployment of AI, assess relevant safety and security initiatives, advise on cyber risks associated with deployment, and identify workforce gaps. It would also brief the relevant congressional committees every six months. These details currently rely on the sponsors' official announcement; the full text in the congressional repository could not be accessed during the research.

Critical infrastructure operators and businesses providing AI services to government should integrate AI security assessments into existing cybersecurity and change management processes. At model integration, record data sources, hosting environments, external service dependencies, and tool access permissions. Before release, test malicious inputs, sensitive information leakage, and unauthorized actions, and define paths for human intervention, permission revocation, and service restoration. During operation, link model versions, configuration changes, monitoring alerts, and incident response records. Procurement terms can also allocate responsibilities for vulnerability notification, security updates, and cooperation with investigations. Existing secure development guidance can inform this preparation and help businesses explain their AI systems' protections in real operating environments. The bill currently proposes institutional coordination arrangements; it does not establish new business obligations for certification, log retention periods, or incident reporting.

A bill with the same name was introduced as H.R. 8348 on 10 May 2024. Records from the previous Congress show introduction and referral to a subcommittee, without enactment. The current bill has been reintroduced in the 119th Congress; committee consideration, text revisions, and possible enactment should be monitored. Sources differ on the date: the legislative record states 6 October 2026, while the sponsors' announcement was published on 8 October 2026. This report uses the legislative record to determine the introduction date. In parallel, CISA, the UK's National Cyber Security Centre (NCSC), and other agencies have already jointly published the advisory Guidelines for Secure AI System Development, covering secure design, development, deployment, and operation and maintenance. The bill would add internal coordination and congressional oversight arrangements to this work. Businesses can continue preparing technical controls under existing guidance while separately tracking the bill's status and eventual scope.

On 6 October 2026, the UK government published its formal response through the Medicines and Healthcare products Regulatory Agency (MHRA), accepting all 44 recommendations made by the National Commission into the Regulation of AI in Healthcare and setting out initial delivery priorities and next steps for healthcare AI regulatory reform.

The response is a government policy paper; acceptance does not mean that every measure is already an obligation in force. Key arrangements include improving the qualification and risk classification of software and AI-enabled medical devices, taking design, functionality, and context of use into account when assessing intended purpose; strengthening post-market surveillance and using Predetermined Change Control Plans (PCCPs) to define the boundaries of permitted iteration; and developing guidance on operational conditions for safe deployment to clarify the responsibilities of manufacturers, healthcare providers, and professionals. Some measures still require consultation, further guidance, or legislation.

Healthcare AI developers and purchasing organizations should translate preparation into controls for versions, operational conditions, and allocation of responsibilities. Before deployment, define the product's intended purpose, intended patient population, and limitations; validate performance in actual clinical workflows; and introduce human review proportionate to risk. Procurement contracts should specify responsibility for update notifications, performance monitoring, reporting of anomalies, and suspension or shutdown. During operation, link product versions, evaluation results, and incident records, with review or suspension triggered by declining performance. The Commission's original report emphasizes that safety depends on the deployment environment, people, and ongoing monitoring, and that contracts should explicitly allocate responsibility for risk controls. These are deployment recommendations informed by the reform agenda; specific requirements still need to be checked against current rules and subsequent documents.

The Commission published its recommendations report on 10 September 2026. The government response moves the recommendations into implementation preparation, with announced milestones including draft guidance on PCCPs by December 2026 and an implementation roadmap covering all recommendations by spring 2027. On the same day as the response, MHRA opened applications for Phase 3 of its AI Airlock regulatory sandbox, focusing on testing post-market surveillance and lifecycle regulation practices. Evidence generated will inform subsequent guidance and policy. The two actions are connected: the policy response sets out the reform programme, while the sandbox tests how continuous monitoring can operate with real products in healthcare settings. Businesses can use these developments to prepare monitoring and change control evidence, but participation in the sandbox or acceptance of the recommendations does not mean that a product has received market authorization.

On 9 October 2026, the European Commission announced a special meeting of the Scientific Panel on artificial intelligence that day to hear expert recommendations on frontier AI safety and security risks, with Executive Vice-President Henna Virkkunen scheduled to attend.

The announcement states that the Scientific Panel had been investigating recent loss-of-control incidents and, together with the European AI Office, had prepared a set of questions for the companies that developed the models involved. The panel was to present recommendations to the Commission during the meeting. It comprises 60 independent experts, with responsibilities including alerting authorities to systemic risks, advising on model classification and evaluation methodologies, and supporting cross-border market surveillance. This is an expert advisory and regulatory support activity. As of 11 October 2026, the research found no publicly available list of the companies involved, question set, or full recommendations. The announcement therefore does not establish that specific penalties have been initiated or new deployment obligations introduced.

Model providers should organize preparation around verifiable safety evidence. Article 55 of the Artificial Intelligence Act already sets out requirements for providers of general-purpose AI models with systemic risk concerning model evaluation, adversarial testing, risk mitigation, documentation and reporting of serious incidents, and cybersecurity. Applicability still needs to be checked against model classification and transitional arrangements. Businesses procuring and deploying models should request evaluation summaries that can be shared, usage limitations, and incident notification arrangements from suppliers. For agents able to invoke code, access networks, or interact with business systems, include tool permissions, execution boundaries, and shutdown capabilities in release acceptance checks. Test unauthorized actions and continued execution of anomalous tasks, and retain linked records of model versions, authorizations, tool calls, and response actions. These are deployment recommendations based on the risk signals; the meeting announcement does not prescribe uniform log retention periods or approval workflows.

The meeting can be understood within the EU's existing framework for frontier model oversight. The Action Plan on Cybersecurity and Artificial Intelligence, published on 7 July 2026, had already proposed strengthening pre-release evaluation and third-party evaluation capacity. This meeting focuses on recent loss-of-control incidents and questions for model developers. At the legal level, Articles 90–92 of the AI Act allow the Scientific Panel to submit a duly reasoned qualified alert and set out procedures for the Commission to request information and conduct model evaluations. Preparing questions must nevertheless be distinguished from issuing a formal statutory information request. Subsequent monitoring should establish whether recommendations are published and whether specific information requests, evaluations, or corrective measures follow; no such subsequent outcomes were found in the current research.

On 8 October 2026, the Office of the Privacy Commissioner of Canada (OPC) updated its Recording of Customer Telephone Calls guidance, clarifying that voiceprints used for identity verification are sensitive biometric information and addressing consent requirements when businesses collect and use that information.

This is an update to guidance under the existing Personal Information Protection and Electronic Documents Act (PIPEDA). In most cases, a customer who continues a call knowing that it is being recorded and why may provide implied consent. Where voiceprints or other sensitive information are involved, express consent may be required. Purpose notices must be specific: “security or quality assurance” cannot serve as a generic description of marketing, customer profiling, or authentication. Alternatives must be offered where recording is not integral to the service and no statutory exception applies. The requirements apply to businesses subject to PIPEDA regardless of who initiates the call; businesses must also ensure that outsourced service providers follow the rules.

Businesses deploying voice-based customer service, call analytics, or voiceprint authentication should integrate purpose and consent checks into data processing workflows. Separately record the purposes of recording, transcription, summarization, voiceprint authentication, and model training, and verify that notices and consent cover actual processing. Where sensitive voiceprints are involved, implement an appropriate express consent process at the extraction stage and test whether processing controls respond correctly when customers refuse or withdraw consent. During procurement, confirm whether suppliers generate voiceprint templates, to whom they disclose data, and whether they support access restrictions and deletion. Set retention periods according to purpose and applicable law, covering cloud storage, backups, and processors acting on the organization's behalf. OPC's biometric guidance already explains that a generic notice stating that recording is for identification purposes is generally insufficient to obtain valid consent. These system configurations are deployment recommendations; this update does not prescribe a uniform technical implementation.

Voiceprint consent requirements have an existing basis in enforcement and guidance. The Rogers investigation findings published on 30 March 2022 determined that the company had failed to obtain valid consent before generating voiceprints. On 11 August 2025, OPC published biometric processing guidance; the latest update applies the relevant interpretation more directly to customer call recording. Regional differences also affect deployment schedules. Quebec requires disclosure of processes using biometrics to verify identity; if a database of biometric characteristics or measurements is created, disclosure must be made at least 60 days before it enters service. Businesses operating nationwide should therefore assess both federal applicability and provincial requirements, incorporating consent workflows, voiceprint database creation, and disclosure milestones into deployment plans.