MI5 chief warns of threats to UK from Russia, Iran and AI. Frank Gardner - Security correspondent; Adam Goldsmith Tue, October 13, 2026 at 5:52 AM PDT Watch: State threats and AI - Key takeaways from MI5 chief's speech MI5 is devoting more resource to countering state threats than during the Cold War, the head of the intelligence service has said, as he labelled Russia "aggressor number one". Sir Ken McCallum warned MI5 was now facing a situation in which "our terrorism and state threats operational missions are pretty much of equal scale" as threats from foreign adversaries have quadrupled over the last decade. He used an annual speech at the domestic intelligence service's headquarters in London to also warn of rising threats from Iran and "extreme right-wing terrorism". Sir Ken added that MI5 had "caught terrorists seeking to exploit AI", including by integrating the technology with drones - and warned the UK must stay ahead of the emerging threats like new bioweapons. The MI5 chief said Russia was waging hybrid warfare while carefully avoiding triggering a joint response from Nato, a line which he said the Kremlin would be "foolish to cross". Sir Ken said Russia had stopped short of "conventional conflict" against the military alliance, and is instead trying to unsettle European countries while its progress in its full-scale invasion of Ukraine is "minimal and horrifically costly". He said MI5's successes in identifying and expelling Russian spies from the UK had left the Kremlin dependent on paying proxies for attempted attacks. He continued: "With its traditional covert footprint eviscerated by almost 1,000 expulsions since the invasion, the Kremlin's global toolkit is weak. So we see them scrambling to pay criminals to do their bidding." He added: "These are not the actions of a great power with the strategic upper hand." Meanwhile, Sir Ken warned that Iran is also expanding a hybrid campaign in the UK. The intelligence chief said there had been 50 arrests in the last year over Iran-linked activities. Much of this Iranian aggression has been aimed at Jewish communities, Sir Ken added, and also at "harder targets, like military sites". Sir Ken did not mention arrests at RAF Fairford directly, where Prime Minister Andy Burnham has said an alleged plot was "almost certainly directed by Iran's Islamic Revolutionary Guard". But the MI5 head continued: "What I can say is that those who seek to cause mayhem on British streets - whether hired thugs here or their paymasters overseas - will meet an unflinching response." Sir Ken added that the National Security Act - a 2023 law designed to target people working against UK interests for foreign powers - is in heavy use, and has been employed to tackle threats from China as well as Iran and Russia. On the emerging threat posed by AI, Sir Ken said: "AI is already supercharging the cybersecurity contest between attackers and defenders. "It can produce propaganda and radicalising material tailored to the individual "We've caught terrorists seeking to exploit Al to improve their reconnaissance and attack methodologies, with Al-enabled drones adding a new twist." The threat he was most worried about, Sir Ken told Thames House, was extremist groups harnessing AI to make bioweapons. He said that threat was "not yet realised" but warned the technology has the "potential to change the stakes on existing or even new bioweapons". Sir Ken continued: "These risks are real. The UK Intelligence Community - MI5, GCHQ, MI6 - is working closely with the UK's globally respected AI Security Institute to take them on." Sir Ken said AI poses unknown threats, telling Thames House: "Put at its simplest, human history has given us no experience of facing an actor which is smarter than we are". In April, the UK's terrorism threat level was raised to severe for the first time in five years, meaning an attack is highly likely. Sir Ken said Islamist groups like Al Qaeda and so-called Islamic State both continue to direct and inspire attacks, while the increased risk was also driven "in significant part" by extreme right-wing terrorism. He continued: "Over the last year the most marked trend has been an increase in cases where extremist versions of anti-immigration and anti-Muslim grievance are tipping into terrorist intent." Sir Ken said extreme right-wing terrorism tended to "skew young", driven by the internet - adding that MI5 had in recent months investigated a "radicalised extreme right-wing 13-year-old here in the UK". He said: "That investigation generated insight that enabled one of our international partners to prevent a lethal attack on their soil."

New framework enhances reproducibility and transparency in AI model evaluation. A collaboration between the UK AI Safety Institute and Hugging Face introduces EvalEval, a framework designed to standardize and improve the transparency of AI model benchmark results, addressing critical challenges in reliable evaluation. Published September 21, 2026 Addressing the reproducibility crisis in AI benchmarking. As artificial intelligence models grow in complexity and scope, the reliability and reproducibility of their evaluations have become a paramount concern. The proliferation of benchmarks and the nuanced variations in evaluation methodologies often lead to inconsistent or incomparable results, hindering true progress and trustworthy assessments. A significant step towards resolving this challenge has been taken through a collaborative effort between the UK AI Safety Institute (UK AISI) and Hugging Face, resulting in the development of EvalEval. This novel framework aims to standardize the reporting and execution of AI model evaluations, providing a robust solution for researchers and developers to ensure their benchmark results are not only accurate but also verifiable and transparent. By fostering a more consistent approach to evaluation, EvalEval is poised to enhance the integrity of AI research and development across the industry. Introducing EvalEval: A unified evaluation ecosystem. EvalEval is designed as a comprehensive system for evaluating large language models (LLMs) and other AI models, emphasizing reproducibility and transparency. It leverages Hugging Face's existing ecosystem, integrating deeply with the Hugging Face Hub, datasets, and models. The core idea is to encapsulate an entire evaluation setup - including the model, evaluation code, data, and environment - within a single, shareable artifact. This artifact, often a repository on the Hub, ensures that anyone can rerun the exact evaluation and obtain identical results. Key features of EvalEval include explicit versioning of all components, from evaluation scripts to the specific model checkpoints used. It supports various evaluation types, from traditional benchmarks to more complex adversarial tests, and provides tools for visualizing and comparing results. This structured approach directly tackles the "reproducibility crisis" in AI, where published results are frequently difficult or impossible to replicate due to undocumented variables or proprietary setups. How EvalEval facilitates transparent benchmarking. At its heart, EvalEval functions by creating a traceable and auditable record of each evaluation. When an evaluation is conducted using the framework, all necessary information is meticulously recorded. This includes the exact commit hash of the model used, the specific version of the evaluation script, the dataset split, and even the computational environment specifications. These details are then packaged and made accessible, typically through a dedicated repository on the Hugging Face Hub. For developers, this means no more guessing which model version yielded which score, or struggling to recreate a testing environment. For researchers, it offers a verifiable path to confirming published results. The framework also supports the creation of "evaluation cards" and "results cards," which provide human-readable summaries of the evaluation methodology and outcomes, further enhancing transparency and ease of understanding for the broader AI community. Impact on AI Safety and development. The UK AI Safety Institute's involvement underscores the critical role of reproducible evaluations in ensuring AI safety. For high-stakes applications, understanding precisely how models perform and the conditions under which those performances were achieved is non-negotiable. EvalEval provides a foundational tool for rigorous safety assessments, allowing for the consistent testing of alignment, bias, and robustness across different models and iterations. Beyond safety, the framework benefits general AI development by accelerating innovation. Developers can more confidently build upon existing work, knowing that benchmark improvements are genuinely representative of enhanced model capabilities, rather than artifacts of evaluation discrepancies. This fosters a more reliable feedback loop for model training and refinement, ultimately leading to more robust and capable AI systems. Why it matters. EvalEval represents a significant leap forward in standardizing AI model evaluation. By making benchmark results transparent, reproducible, and verifiable, it addresses a core challenge that has long plagued AI research and development. This initiative will not only bolster confidence in reported AI performance metrics but also provide a crucial foundation for safer, more robust, and continually improving artificial intelligence systems across all sectors.