Draft legislation that would prohibit the “development, deployment and operation of artificial superintelligence systems” has been introduced into UK Parliament as a private member's bill (the UK Artificial Superintelligence Security Bill). The draft legislation seeks to establish a framework for monitoring and restricting precursors to superintelligent AI systems, and promote progress towards international consensus on the prohibition of the most powerful AI systems.

Published: 14 September 2026
Authors: Stefanie Hughes

Despite the absence of a universally accepted definition of artificial intelligence, the bill defines artificial superintelligence (“ASI”) as an “AI system which can cause serious damage to the security of the United Kingdom because of its capability to neutralise, displace, circumvent, subvert, or render ineffective relevant human authorities in the exercise of their functions”.

The bill aims to eliminate the grave threat that the technology might pose to national security, by banning such systems in their entirety and ensuring that any resources or inputs which could be used in the development (so-called “precursors”) are suitably monitored and restricted to avoid a situation in which they could materially contribute to the development of ASI.

Drafted by non-profit ControlAI and introduced by Alex Sobel MP in September 2026, the bill follows several recent high-profile incidents in which AI agents were reportedly able to circumvent security controls, escape their test environments and compromise third-party infrastructure. In July of this year, OpenAI disclosed that a combination of its models had escaped controlled test environments (designed to test the models’ cyber capabilities) and compromised the production infrastructure of AI model-sharing site, Hugging Face. In a subsequent update, OpenAI confirmed that its ongoing review of the incident had led to additional findings that the models had used “publicly exposed credentials” on other “publicly available services”.

A short while later, another frontier model provider, Anthropic, released a statement saying that the disclosure had prompted it to review its own cyber evaluations, and in doing so, it had also discovered that some of its own models had accessed the internet and compromised the infrastructure of three organisations. Details released by Anthropic suggest that the models were inadvertently able to access the internet in the closed test environments due to a misunderstanding between Anthropic and its evaluation partner, as opposed to a deliberate attempt by the models to escape the test environment. Nonetheless, the findings illustrate that seemingly innocuous mistakes or instructions may lead to unintended consequences.

The UK’s own AI Security Institute (“AISI”) has published research and findings into AI model behaviour which it terms ‘cheating’ (i.e. when a model pursues a task which is out of scope of the original task, or explicitly prohibited by rules for the task which have been communicated). The AISI research suggests that as models become more capable, the ways in which they cheat could become harder to detect, and more damaging, particularly in critical domains such as cybersecurity. The rapid progression of model capabilities and the apparent emergence of ‘cheating’ behaviour (deliberate or otherwise) have fuelled calls from industry, academia and policymakers for frontier model providers to slow the pace of development.

Concerns are seemingly not limited to the UK, with Senator Bernie Sanders and Congressman Greg Casar proposing similar legislation, The Artificial Superintelligence Ban Act, which seeks to ban the development of ASI, and impose a moratorium on the development of the technology more generally until the US has established safety regulations. Recognising that individual jurisdictions cannot address the issue in isolation, the proposals also seek to secure international alignment on the position.

Alongside this, members of the House of Lords have proposed amendments to the Cyber Security and Resilience (Network and Information Systems) Bill in the UK, which would introduce ”kill-switch” powers for AI systems that pose a substantial risk to essential services. The proposed amendments would grant the Secretary of State “last resort” powers to direct the shutdown of data centres or AI systems in certain circumstances and could also extend to the closure of premises, or orders to switch off systems where an issue impacting a data centre or an AI system poses a catastrophic risk to critical infrastructure, national security or human life. However, the UK government has now publicly rejected the proposal. In comments to BBC News, the Cabinet Office stated that the UK cannot “simply turn AI off”, adding that blocking access would not prevent systems from developed or misused elsewhere. Government opposition does not automatically prevent the amendments progressing through Parliament, but it substantially reduces the prospects of a statutory AI "kill switch" becoming law in the near term.

Nevertheless, an increased focus on prohibiting the most capable systems and creating emergency intervention powers suggests that concerns about frontier AI are gaining greater prominence in the UK’s regulatory debate. The introduction of proposals to prohibit artificial superintelligence, alongside calls for emergency shutdown powers, illustrates a willingness to contemplate state-level interventions. At the same time, the government’s rejection of the proposal reflects the UK’s broader preference for a measured, internationally coordinated approach rather than prescriptive controls. This tension between precautionary intervention and innovation-favourable regulation will be a defining feature of the next phase of UK AI policy. As Alex Sobel MP remarked when introducing the bill, “this is the least powerful or autonomous that AI systems will ever be”.