正文 · AI 翻译
IT Home reported on October 12 that, according to The Guardian, the UK will strengthen its cyber defenses while calling on leading artificial intelligence developers to take more proactive steps to govern the security of their systems.
UK Minister of State for AI Kanishka Narayan is expected to formally announce these policy positions in a keynote speech in west London.
Narayan will call on frontier AI developers to take more concrete steps to keep their systems secure. IT Home has learned that Narayan said the UK must never rely entirely on tech giants to “self-regulate” their models, and stressed that the UK must build more AI infrastructure under its own control.
Further reading
According to information on the UK government website, Kanishka Narayan was appointed Minister of State for AI on July 20, 2026, with responsibilities in both the Cabinet Office and the Department for Business, Innovation, Science and Trade (BIST). He previously served as Parliamentary Under-Secretary of State in the Department for Science, Innovation and Technology (DSIT) from September 7, 2025, to July 20, 2026, and was elected MP for the Vale of Glamorgan in Wales in July 2024.
Narayan’s responsibilities in the Cabinet Office include the AI Security Institute (AISI). The institute aims to build an evidence base on AI risks to inform government decisions. According to information disclosed by the UK Parliament in April 2026, AISI has worked with leading AI companies, including OpenAI, Google DeepMind and Anthropic, to test more than 30 models, and its findings have led to practical changes before the models were deployed.
On December 18, 2025, AISI released its first Frontier AI Trends Report. The report showed that AI systems’ capabilities were improving rapidly: in cybersecurity, models’ success rate on apprentice-level tasks rose from less than 9% in 2023 to about 50% in 2025, and in 2025, a model completed an expert-level cyber task requiring up to 10 years of experience for the first time; in software engineering, models’ success rate on programming tasks taking about 1 hour rose from less than 5% 2 years earlier to more than 40%; and for some cyber tasks, the length of tasks AI can complete without human guidance roughly doubles every 8 months. The report also noted that safeguards intended to ensure AI behaves as expected are improving: the time it takes red teams to find a “universal jailbreak” has increased from minutes to hours, an improvement of about 40-fold.
The UK government has previously introduced a series of measures to strengthen domestic AI infrastructure: it established a £500 million Sovereign AI venture capital fund, whose first equity investment went to AI infrastructure company Callosum; released an AI Hardware Plan to support UK companies developing AI chips and semiconductor technology; published a Cyber Growth Action Plan to boost the domestic cybersecurity industry; and expanded the National Security Strategic Investment Fund. The UK government believes that the vast majority of AI systems should be regulated at the point of use.