The UK gives ministers power to direct or halt a frontier lab

Where things stand (11 Oct 2026)

No qualifying Act yet. The May 2026 King's Speech had no standalone AI bill (GOV.UK, 13 May 2026; SecurityBrief, 14 May). Nothing relevant has received Royal Assent since the question opened on 16 Sept. The National Security (State Threats) Act received Royal Assent on 8 Jul 2026, before the window.

The live vehicle is the Cyber Security and Resilience (Network and Information Systems) Bill (bill 4035). I checked it on the Parliament bills API. It is a carried-over government bill at Lords Report stage, which is set for 26 Oct 2026. Its existing section 43 lets the Secretary of State direct "regulated persons" for national-security reasons. That is a general power, so under the question's rules it does not count without an AI-specific addition.

Amendments on the table:

• The Lords Report running list (9 Oct) includes Lord Clement-Jones's "last-resort powers". These are regulations that could give the Secretary of State power to order the shutdown of data centres or "AI systems deployed on a substantial scale" in an AI emergency.
• Lord Tarassenko proposes a pre-deployment check by the AI Security Institute. The Secretary of State could then block, restrict or suspend a frontier model.
• Baroness Harding proposes incident reporting for frontier developers.
• In Grand Committee on 1 Sept (Hansard), the Conservative front bench (Viscount Camrose) "support[ed] the principle" of the shutdown power and wanted powers held by "accountable" ministers rather than the AI Security Institute. Minister Baroness Lloyd rejected the amendment but said the government was examining "proportionate containment powers … including powers to restrict access to specific AI systems".

The government has since moved. On 9–10 Oct, The Times reported that ministers will table amendments to this bill "this month" to stop AI "loss of control" and to require incident reporting. It called this Britain's "first AI safety legislation" (summary: Digital Watch). Other signals:

• The PM told the UN on 22 Sept that "fundamental decisions" on AI "must be taken by elected governments", and that new laws would come if needed.
• The Cabinet Office earlier said "Britain cannot simply turn AI off" (PA/Irish News).
• The AI minister's 7 Sept statement framed containment mainly as security practice.

The background is unusually alarming: rogue-agent incidents, Anthropic and OpenAI calling for curbs (Guardian, 14 Sept), and more than 70 MPs and peers pressing. Alex Sobel's Artificial Superintelligence Bill is a Ten Minute Rule bill, which is close to zero chance.

Model

Reference class. UK governments facing Lords defeats usually concede reviews rather than substance, as with Lady Kidron's AI and copyright amendments to the Data (Use and Access) Act 2025. Here, though, the government is pre-empting the Lords with its own AI amendments. UK national-security powers are normally given to the Secretary of State.

Path A: this bill. I put P(enacted with qualifying text, accepted by the resolver) at about 0.35:

• About 0.29 from the government's own amendments. I estimate 30% that they include an explicit containment, restriction or suspension direction, plus partial credit for regulation-making or regulated-person-only variants.
• About 0.06 from the Lords passing the shutdown power and the government conceding a version during ping-pong.

Timing of Royal Assent:

• The Commons Christmas recess starts 17 Dec, and the session runs to at least July 2027.
• I put P(Royal Assent by 31 Dec) at about 0.55 and by 31 Mar 2027 at about 0.9.

Path B: a later law (2027–31). This could be a frontier-AI or AI-security bill (for example mandatory testing with a ministerial power to block deployment), implementation of the planned G20 AI principles, or emergency legislation. I used quarterly hazards of 0.3–2.5%, lower around the 2029 election, giving about 0.29 cumulative.

Combined. I computed P = 1 − (1 − P_A(t))(1 − P_B(t)) in Python. That gives 0.18 at end-2026, 0.32 at end-Q1 2027 and 0.54 by end-2031.

Main risks to the estimate

Against YES. The government's amendments may contain only incident reporting and containment duties enforced by regulators. The Times headline does not mention shutdown powers, and the government has repeatedly dismissed kill switches. US hostility to rules on US labs, the government's focus on growth, a possible change of government in 2029, or a preference for an independent regulator could also block later ministerial powers.

For a higher number. The minister herself floated restricting access to specific AI systems. The Lords majority (Conservatives 247 + Lib Dems 80 against Labour's 233) can force the issue, and public opinion strongly favours government power to suspend harmful AI. The published amendment text, expected within days, is the key update.

Question. By the last day of each quarter, has the UK Parliament enacted a law specifically empowering UK ministers to direct, halt, or take control of a covered developer’s UK frontier-AI research, training, or deployment operations?

Definitions.

• Enacted law: an Act of the UK Parliament that has received Royal Assent. An amending provision in a broader Act counts. Commencement of the power and its actual exercise are not required; bills, delegated legislation alone, and Acts of devolved legislatures do not count.
• UK ministers: a UK Government minister, including a Secretary of State, acting directly or through a person acting on the minister’s behalf. A power vested solely in an independent regulator or court does not count unless the Act also empowers a minister to require its exercise in a particular case.
• Specifically empowering: the enacted text expressly makes AI development or deployment, a defined advanced-AI category, or the frontier-AI operations of a named developer subject to the relevant ministerial power. General emergency, national-security, investment-screening, competition, or corporate powers that could incidentally reach AI operations do not count without an enacted AI-specific extension.
• Frontier AI: general-purpose AI models at or approaching the leading edge of capabilities when developed, including systems intended to surpass that edge. A statutory category of advanced, frontier, systemic-risk, or superintelligent AI counts if it covers such models; a law covering AI more broadly also counts if its operational powers cover these models.
• Covered developer’s UK operations: research or training physically conducted in the UK, or deployment through making a model available for use in the UK, by a covered developer or a controlled entity conducting that developer’s AI operations. The Act need not name a developer or establish that a particular model already meets a future capability threshold; its jurisdiction and regulated class must permit application to at least one covered developer’s qualifying operations.
• Direct, halt, or take control: any one of (i) issuing legally binding operational instructions about whether or how qualifying research, training, or deployment is conducted; (ii) ordering its suspension, termination, or noncommencement, including withholding or withdrawing legally necessary ministerial permission; or (iii) assuming or appointing someone to assume operational management or control. Conditional powers, including powers dependent on an emergency declaration, risk finding, or statutory procedure, count.

Resolves YES if an enacted Act published within the resolution window grants at least one of these powers over at least one of the three specified activities. A power over a single project or model suffices; authority over the whole developer is not required.

Exclusions (resolve NO).

• Proposed legislation, consultations, voluntary agreements, ministerial announcements, or nonbinding guidance.
• Reporting, evaluation, inspection, model-access, information-sharing, or fine-imposition powers alone.
• A statutory prohibition enforceable only by courts or an independent regulator, without a qualifying ministerial operational power.
• Funding or procurement conditions alone, or powers limited to downstream users rather than a covered developer’s operations.
• Existing general powers, including section 26 of the National Security and Investment Act 2021, without a qualifying new AI-specific enactment. Republishing or invoking an older Act does not qualify.

Resolving source. The authoritative enacted text on legislation.gov.uk, with Royal Assent and publication corroborated where necessary by UK Parliament’s bill record at bills.parliament.uk or the official parliamentary record at hansard.parliament.uk. These legislative sources, rather than developer confirmation under A2, control resolution. Apply the resolution window and A3 publication-date convention: use the first official publication establishing both enactment and the qualifying text, not a bill’s earlier publication date or the power’s commencement date.