UK Lords Debate Whether Frontier AI Developers Must Report When Their Models Breach British Networks

The Cyber Security and Resilience Bill began report stage in the House of Lords on Monday, its last line-by-line opportunity for amendment, and the amendment paper shows peers trying to bolt the UK's first statutory AI duties onto a cybersecurity vehicle. Day-one amendments up for debate include the reporting of incidents involving frontier AI models and the voluntary notification of incidents, cyber threats and near misses, according to the UK Parliament news service. Watch on Parliament TV began at 3:15pm London time.

The most consequential amendment comes from Baroness Dido Harding, who wants developers of the most capable AI models to tell regulators when their models are used to break into a regulated network, with penalties for failing to report. That shifts the duty from the victim to the toolmaker: the bill as passed by the Commons obliges the regulated operator to report an incident, while Harding's clause treats a frontier model as part of the attack chain, a category the legislation never contemplated.

Two other amendments would build enforcement muscle. Lord Tim Clement-Jones proposes regulations giving the Secretary of State last-resort powers to order data centres or large AI systems shut down during an AI security emergency, alongside a narrowing of the "managed service" definition to providers with privileged access, which decides which IT suppliers fall inside the law. Lord Lionel Tarassenko wants the AI Security Institute to audit frontier AI products before they can be used in regulated systems, which would hand AISI a regulatory role for the first time. Baroness Ruth Kidron has tabled a broader prohibition on making available in the UK any AI product with nine listed capabilities, including evading human shutdown, enforced through a pre-launch safety case.

The bill underneath these amendments is substantial on its own. It strengthens cyber duties for providers of essential services such as healthcare, water and energy, brings data centres and managed service providers into scope, and updates incident reporting to an initial notice within 24 hours and a full report within 72. It cleared the Commons on June 16, completed second reading in the Lords on July 14, went through Grand Committee in early September, and received Scottish legislative consent on October 7. Royal Assent is expected between late 2026 and spring 2027, with most duties switched on by secondary legislation around 2028.

The political backdrop is moving the same way. The UK government said this month it is considering how the bill could cover AI near-miss reporting and loss-of-control risks, and the Financial Conduct Authority's chief data officer Jessica Rusu told a Bloomberg audience that existing incident-reporting rules "must extend to significant AI failures and near-misses." Critics see two traps: without penalties the duty becomes a paper tiger, and a near-miss obligation can reward the firms with the weakest monitoring, since they will find fewer events to report.

The structural problem is that AISI, the body every amendment leans on, is a directorate of government rather than a statutory entity, with no board, no power to compel a model and no appeal process. The Joint Committee on Human Rights asked in September for it to be put on a statutory footing, and the point is not academic: when AISI was shut out of a frontier model launch in September, restoring access took head-of-state-level diplomacy. Giving a courtesy a legal name is a different thing from giving it a legal basis.

The realistic odds are the usual ones. Most backbench amendments are withdrawn or voted down unless the government backs them, and the precedent is sobering: Baroness Kidron's AI copyright amendments to the Data Use and Access Bill passed back and forth between the Houses five times in 2025 before peers gave way and the bill became law without them. But if any of these clauses survives, the UK would become the first jurisdiction outside the EU AI Act to put a statutory incident-reporting duty on frontier model developers, and it would have done so in a cybersecurity bill rather than an AI one.