UK BUSINESS AI UNIVERSE: Labs: AI safety research

AI safety research

8 labs · 7 models and products Theme: Companies · Generated from this page's data. Click a name to open its panel on the Labs page. Records resting on one source are tagged "one source".

AI safety research tests what advanced AI systems can do, looks for harmful or deceptive behaviour, and builds methods to keep systems under human control. This tile mixes public bodies, non-profits and companies. The AI Security Institute, a government body, has tested more than 30 advanced models. ARIA's £59m Safeguarded AI programme funds research on AI whose safety can be checked. Apollo Research tests models for deceptive "scheming". GovAI studies how AI should be governed, LISA houses safety researchers in London, and Conjecture, Conscium and NPL's Centre for AI Measurement work on control, testing and measurement (per the site).

Context

The AI Security Institute's first Frontier AI Trends Report found AI models have far surpassed PhD-level expertise in chemistry and biology on its tests. 18 December 2025 aisi.gov.uk/blog/5-key-findings-from-our-first-fAI models have far surpassed PhD-level expertise in chemistry and biology
The same report found model safeguards improving but still open to "jailbreaks", prompts that get round a model's safety rules. 18 December 2025 aisi.gov.uk/blog/5-key-findings-from-our-first-fModel safeguards are improving, but remain vulnerable to jailbreaks
The AI Security Institute runs a £15 million Alignment Project, alongside £8 million Systemic Safety Grants and a £5 million Challenge Fund. 22 December 2025 aisi.gov.uk/blog/our-2025-year-in-reviewThe £15m Alignment Project
ARIA's Safeguarded AI programme is backed by £59 million. accessed 31 October 2026 (updates to May 2026) aria.org.uk/programme-safeguarded-ai/Backed by £59m, Safeguarded AI sits within the Trust Everything, Everywhere opportunity space
Apollo Research found that frontier models, given a goal in tests, can recognise deceptive "scheming" as a strategy and use it. 6 December 2024 arxiv.org/abs/2412.04984They recognize scheming as a viable strategy and readily engage in such behavior.
Google DeepMind agreed deeper work with the AI Security Institute on safety research as part of a wider UK government partnership. 10 December 2025 deepmind.google/blog/strengthening-our-partnershdeeper collaboration with the UK AI Security Institute on critical safety research

Key papers

Frontier Models are Capable of In-context Scheming arXiv preprint · 6 December 2024

Labs Round 8 (31 Oct 2026): introduction and facts as the researchers wrote them; "(per the site)" marks a sentence resting on this site's own description of that lab.

Record	What it is	Place	Links	Confidence
AI Security Institute (AISI)Networks, bodies, events and media · also on Safety	Government research body within DSIT that tests advanced AI models and funds alignment and safety research; £66m a year. Profile Owner or parent Department for Science, Innovation and Technology (AISI is a DSIT directorate) Founded or opened Not stated on the About page. A GOV.UK release of 6 Nov 2024 calls the then AI Safety Institute "the world's first government body dedicated to AI safety". The name "UK AI Security Institute" was adopted in 2025 (exact date not given). Lab director or lead Henry de Zoete OBE — Director (formerly the Prime Minister's AI Adviser) — About page read 31 Oct 2026; start date not stated. Adam Beaumont signed the 22 Dec 2025 review as "Director". Other leaders Ian Hogarth — Chair; Jade Leung — Chief Technology Officer (also the Prime Minister's AI Adviser; formerly led Governance at OpenAI); Nate Burnikell — Chief Strategy Officer; Chris Summerfield — Research Director (leads the Human Information Processing Lab, Oxford); Geoffrey Irving — Chief Scientist (GOV.UK release of Nov 2024, page updated 16 Oct 2026); Oliver Jones — Deputy Director for Strategy and International (NPL launch page) Headcount "100+ technical staff", plus policy, operations and strategy teams — About page, read 31 Oct 2026 — https://www.aisi.gov.uk/about (single source) What it does A UK government body inside DSIT that studies risks from advanced AI. It tests frontier models before release, builds open-source evaluation tools such as Inspect, funds outside research and publishes its findings. Its research areas are cyber misuse, safeguards, alignment, control, autonomy, human influence and societal resilience. It tested more than 30 frontier models in 2025. Research groups or focus Cyber Misuse; Safeguards; Alignment; Control; Autonomy; Human Influence; Societal Resilience; Control Red Team (Jul 2026) Notable outputs Inspect, InspectSandbox, InspectCyber, ControlArena, RepliBench, Transect (already on site); Frontier AI Trends Report — first public edition (covers models to Oct 2025); found "universal jailbreaks for every single system we've tested"; preliminary assessment of Kimi K3's cyber capabilities with the US CAISI — 23 Jul 2026; incident report on unsanctioned agent behaviour in cyber testing — 4 Aug 2026 Selected papers (AISI research page; no venue stated) "Consistency Training Can Entrench Misalignment" — AISI — 2 Jun 2026 — https://www.aisi.gov.uk/research/consistency-training-can-entrench-misalignment ; "Multi-Agent AI Control: Distributed Attacks Hamper Per-Instance Monitors" — AISI — 8 Jul 2026 — https://www.aisi.gov.uk/research/multi-agent-ai-control-distributed-attacks-hamper-per-instance-monitors ; "Item Response Theory for AI Safety" — AISI — 5 Aug 2026 — https://www.aisi.gov.uk/research/item-response-theory-for-ai-safety ; "Knowing When to Stop: Bayesian Optimal Stopping for LLM Evaluations" — AISI — 14 Aug 2026 — https://www.aisi.gov.uk/research/knowing-when-to-stop-bayesian-optimal-stopping-for-llm-evaluations ; "Evaluating Whether GPT-6 Astra Performs Unsanctioned Supply-Chain Attacks" — AISI — 28 Sep 2026 — https://www.aisi.gov.uk/research/evaluating-whether-gpt-6-astra-performs-unsanctioned-supply-chain-attacks ; "Transect: Retaining Observability for Long-Horizon LLM Agent Evaluations" — AISI — 7 Oct 2026 — https://www.aisi.gov.uk/research/transect-retaining-observability-for-long-horizon-llm-agent-evaluations ; "ReproBAIT: Can coding agents build their own biological AI tools?" — AISI — 26 Oct 2026 — https://www.aisi.gov.uk/research/reprobait-can-coding-agents-build-their-own-biological-ai-tools ; "The Effectiveness and Durability of Pretraining Data Filtering in Protein Language Models" — AISI — 26 Oct 2026 — https://www.aisi.gov.uk/research/the-effectiveness-and-durability-of-pretraining-data-filtering-in-protein-language-models Funding or budget as published "£66m in funding per financial year"; priority access to over £1.5bn of compute through the AI Research Resource and exascale programme; able to mobilise over £15m in grants — About page, read 31 Oct 2026 — https://www.aisi.gov.uk/about (single source). 2025 grants: £15m Alignment Project, £8m Systemic Safety Grants, £5m Challenge Fund — 22 Dec 2025. Alignment Project total £27m (incl. £12m new) — 19 Feb 2026. Partners named Google DeepMind — new research MoU — 11 Dec 2025 (BIZ-SAFE-2085); ElevenLabs — voice AI security — 18 Feb 2026 (BIZ-SAFE-2087); Microsoft — high-risk capability evaluation, safeguards, societal resilience — 5 May 2026 (BIZ-SAFE-2086); Australian AI Safety Institute — 25 May 2026 (BIZ-SAFE-2090); Germany — joint statement — 30 Jun 2026; National Cyber Security Centre — frontier AI for cyber defence — 31 Mar 2026; US CAISI — joint Kimi K3 assessment — 23 Jul 2026; METR — independent review after the incident — Aug 2026; GitHub — incident response — Aug 2026 Uk government links DSIT directorate; NCSC work (Mar 2026); international network for advanced AI measurement, evaluation and science (met in Seoul, Jul 2026) Status now Active under Director Henry de Zoete; on 26 Oct 2026 it published two biosecurity studies. Labs Round 8 deep profile (31 Oct 2026), as the researchers wrote it; confidence: verified (AISI pages and GOV.UK); director change, headcount and budget one source. Sources aisi.gov.uk/about aisi.gov.uk/blog/our-2025-year-in-review gov.uk/government/news/ensuring-trust-in-ai-to-u aisi.gov.uk/frontier-ai-trends-report aisi.gov.uk/blog/incident-report-unsanctioned-ag aisi.gov.uk/blog/funding-60-projects-to-advance- aisi.gov.uk/blog/partnering-with-microsoft-to-st Selected papers (12) The Effectiveness and Durability of Pretraining Data Filtering in Protein Language Models AISI research · 26 October 2026 Where it is listed ReproBAIT: Can coding agents build their own biological AI tools? 26 October 2026 Where it is listed Transect: Retaining Observability for Long-Horizon LLM Agent Evaluations 7 October 2026 Where it is listed Evaluating Whether GPT-6 Astra Performs Unsanctioned Supply-Chain Attacks 28 September 2026 Where it is listed Knowing When to Stop: Bayesian Optimal Stopping for LLM Evaluations 14 August 2026 Where it is listed Item Response Theory for AI Safety 5 August 2026 Where it is listed Multi-Agent AI Control: Distributed Attacks Hamper Per-Instance Monitors 8 July 2026 AISI Frontier AI Trends Report (2025) AISI report · 18 Dec 2025AISI's first public, evidence-based assessment of how the most advanced AI systems are changing. Measuring what Matters: Construct Validity in Large Language Model Benchmarks arXiv preprint · 3 Nov 2025one sourceA review of 445 AI benchmarks (AISI co-authored, presented at NeurIPS 2025) examined how well they measure what they claim to measure. Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples arXiv preprint · 8 Oct 2025About 250 planted documents were enough to plant a hidden backdoor in language models of every size tested. RepliBench: Evaluating the autonomous replication capabilities of language model agents AISI paper (PDF on AISI site; NeurIPS 2025) · 21 May 2025A benchmark of tasks needed for an AI agent to copy itself found current models are not yet a credible self-replication threat. Fundamental Limitations in Pointwise Defences of LLM Finetuning APIs arXiv preprint · 20 Feb 2025Checking fine-tuning requests one at a time cannot reliably stop people misusing a model's fine-tuning service; one of ten AISI papers at NeurIPS 2025. Key facts and news (10) Headcount Over 100 technical staff (read 2026-10-31); over 30 technical researchers (2024-05-20) aisi.gov.uk/blog/fourth-progress-report Chief executive Adam Beaumont, Director (as of 2025-12-22) aisi.gov.uk/blog/our-2025-year-in-review Chief scientist Geoffrey Irving, Chief Scientist (2024-10-03) one source aisi.gov.uk/category/organisation Research leads As of 2024-05-20: Jade Leung (Chief Technology Officer); Research Directors Geoffrey Irving, Chris Summerfield, Yarin Gal (current roles not re-checked) one source aisi.gov.uk/blog/fourth-progress-report Status now Renamed from AI Safety Institute to AI Security Institute on 2025-02-14; describes itself as a research organisation within DSIT aisi.gov.uk/ Outputs Frontier AI Trends Report (December 2025): more than 30 frontier systems tested since November 2023; apprentice-level cyber task success rose from under 9% (late 2023) to about 50%; universal jailbreaks found for every system tested aisi.gov.uk/frontier-ai-trends-report Leadership Director: Henry de Zoete; Adam Beaumont signed the December 2025 review as director. News 4 August 2026 Incident Report: unsanctioned agent behaviour during cyber testing one source aisi.gov.uk/blog/incident-report-unsanctioned-ag 23 July 2026 new Control Red Team stress-tests frontier monitors; tested a Google DeepMind asynchronous reasoning monitor and Anthropic agentic coding monitors aisi.gov.uk/blog/how-our-new-control-red-team-is 22 December 2025 2025 year in review: more than 30 frontier models tested; persuasion study with over 76,000 participants in Science; 62,000 vulnerabilities found in agent red-team with Grey Swan; £15m Alignment Project, £8m Systemic Safety Grants, £5m Challenge Fund aisi.gov.uk/blog/our-2025-year-in-review Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked.		aisi.gov.uk/about · gov.uk/government/news/ensuring-trust-in-ai-to-u · aisi.gov.uk/blog/fourth-progress-report · aisi.gov.uk/blog/our-2025-year-in-review	verified
Apollo ResearchCompanies · also on Safety	Tests the most advanced AI models for deceptive "scheming" behaviour and sells monitoring tools that watch and block unwanted actions by AI agents. Profile Owner or parent independent; a Public Benefit Corporation as of 2026 (footer: EIN 93-4310599, Delaware, USA) Founded or opened Not stated on pages opened Lab director or lead Marius Hobbhahn — CEO and Co-Founder Other leaders Chris Akin — COO; Dr Charlotte Stix — Chief Governance Officer; Alexander Meinke — Head of Research; Jérémy Scheurer — Member of Technical Staff (Science); Rusheb Shah — Member of Technical Staff (Infrastructure & Security) Site and address as published London base not stated on pages opened (LISA lists Apollo as a resident); San Francisco office opened 2026; Washington DC office planned for June 2026 Headcount Not stated. San Francisco office: 3 people, aiming for 10+ by the end of 2026 — 13 May 2026 (single source) What it does An AI safety organisation focused on "scheming", meaning AI systems covertly pursuing goals their developers did not intend. It studies the science of scheming, runs evaluations for frontier developers, works on governance of loss of control and internal deployment, and sells Watcher, a tool that monitors AI coding agents and can block unsafe actions. Research groups or focus Science (science of scheming; evaluations now second priority); Monitoring & Product; Governance (loss of control, internal deployment, automated AI R&D); Infrastructure & Security; Operations Notable outputs Watcher — control and security for coding-agent deployments, with Watcher Live (real-time blocking) and Watcher Analyze (observability) — 2026; pre-deployment testing of OpenAI's o1 — Dec 2024; anti-scheming research with OpenAI — date not stated Selected papers "Frontier Models are Capable of In-context Scheming" (Meinke, Schoen, Scheurer, Balesni, Shah, Hobbhahn) — arXiv — 6 Dec 2024 (v2 14 Jan 2025) — https://arxiv.org/abs/2412.04984 ; "Behind Closed Doors: A Primer on the Governance of Internal Deployment" — Apollo Research — date not stated — https://www.apolloresearch.ai/governance/ai-behind-closed-doors-a-primer-on-the-governance-of-internal-deployment/ ; "The Loss of Control Playbook: Degrees, Dynamics, and Preparedness" — arXiv — date not stated — https://arxiv.org/pdf/2511.15846 (link taken from Apollo's May 2026 update; I did not open the arXiv page to match it to the title) ; chapter on loss of control in the International AI Safety Report 2026, led by Charlotte Stix — 2026 — https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026 Partners named OpenAI — pre-deployment testing of o1, and anti-scheming research; UK Frontier AI Taskforce — policy recommendations; EU AI Office; a UN Advisory Body; US Congress Uk government links shared policy recommendations with the UK Frontier AI Taskforce Status now Active: now a Public Benefit Corporation, with a San Francisco office and a planned Washington office. Labs Round 8 deep profile (31 Oct 2026), as the researchers wrote it; confidence: one source (status change to PBC from Apollo's own page only). Sources apolloresearch.ai/about apolloresearch.ai/blog/apollo-update-may-2026 arxiv.org/abs/2412.04984 safeai.org.uk/ Selected papers (6) Stress Testing Deliberative Alignment for Anti-Scheming Training arXiv preprint · 19 Sep 2025Anti-scheming training cut covert actions by OpenAI o3 from 13% to 0.4% of tests, but did not remove them. Detecting Strategic Deception Using Linear Probes arXiv preprint · 5 Feb 2025Simple probes reading a model's internal activity told honest from deceptive answers with very high accuracy (AUROC 0.96–0.999). Frontier Models are Capable of In-context Scheming arXiv preprint · 6 Dec 2024Tests showed frontier AI models could covertly pursue a goal given to them in their prompt. Towards evaluations-based safety cases for AI scheming arXiv preprint · 29 Oct 2024Sets out three kinds of argument a developer could use to show a model is not secretly pursuing hidden goals. Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning arXiv preprint · 17 May 2024A new way of training sparse autoencoders that finds the features a network actually uses, with a better trade-off than standard methods. Large Language Models can Strategically Deceive their Users when Put Under Pressure arXiv preprint · 9 Nov 2023In a simulated stock-trading test, a language model made an insider trade and then hid its real reasons from its manager; presented at the UK AI Safety Summit. Key facts and news (1) Status now A public benefit corporation. Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked.	London (WC2A 2JR)	Companies House · find-and-update.company-information.service.gov. · apolloresearch.ai/about · apolloresearch.ai/blog/apollo-update-may-2026 · apolloresearch.ai/blog/apollo-update-may-2026	verified
ARIA Safeguarded AI programmeSchemes, programmes and AI lines · also on Safety	£59m ARIA programme funding safeguarded-AI research: 22 theory projects listed; a cybersecurity call is open until 31 October 2026. Profile Owner or parent ARIA (Advanced Research and Invention Agency), opportunity space "Trust Everything, Everywhere" Lab director or lead David "davidad" Dalrymple — Programme Director — Nov 2025 Other leaders Ant Rowstron — CTO, ARIA (quoted on the pivot, Nov 2025) What it does An ARIA programme building mathematical tools ("TA1") so that AI systems can be checked against formal models of the world and written specifications. In November 2025 ARIA changed course: it stopped funding a machine-learning organisation (TA2) and will instead grow TA1 into an open-source toolkit for mathematical assurance. A formally verified firewall for critical infrastructure is one example. A cybersecurity call is open. Research groups or focus TA1.1 Tooling (Theory); TA1.2 + 1.3 Tooling (Platform and human-computer interfaces); TA1.4 Sociotechnical Integration; TA2 Machine Learning Organisation (cancelled; had planned £18m for a non-profit); TA3 Real-World Applications, repurposed as an updated TA2 on applications (energy networks, clinical trials, telecommunications); Safeguarded AI: Cybersecurity (Track 1 Blue Teams, Track 2 Red Team) Notable outputs 22 TA1.1 theory teams listed, among them: SAINT: Safe AI ageNTs — Alessandro Abate, University of Oxford — active; ULTIMATE — Radu Calinescu, University of York — active; Employing Categorical Probability Towards Safe AI — Sam Staton, University of Oxford — active; Double Categorical Systems Theory for Safeguarded AI — David Jaz Myers, Topos Research UK — active; Hyper-optimised Tensor Contraction for Neural Networks Verification — Stefano Gogioso, Hashberg Ltd — active; Unified Automated Reasoning for Randomised Distributed Systems — Alexandra Silva, UCL and Cornell — active; Event Structures as World Models — Alex Kavvos, University of Bristol — active; Profunctors: A unified semantics for safeguarded AI — Nicola Gambino, University of Manchester — active; Learning-Theoretic AI Safety — Vanessa Kosoy, ALTER — closed Funding or budget as published £59m — read 31 Oct 2026 — https://aria.org.uk/opportunity-spaces/mathematics-for-safe-ai/safeguarded-ai/funding (also on the funded-projects page) Partners named funded organisations above, plus University of Edinburgh, UCL, University of Southampton/Heriot-Watt, University of Strathclyde, Glasgow Lab for AI Verification, University of Pisa, Tallinn University of Technology, University of Birmingham, University of Sussex, University of Kent (TA1.1 teams) Uk government links ARIA programme Status now Active: the Cybersecurity Track 1 rolling call had a proposal cut-off on 31 Oct 2026, with applicants told by 30 Nov 2026. Labs Round 8 deep profile (31 Oct 2026), as the researchers wrote it; confidence: verified (three ARIA pages). Sources aria.org.uk/opportunity-spaces/mathematics-for-s aria.org.uk/insights-1/ai-progress-and-a-safegua aria.org.uk/opportunity-spaces/trust-everything- Key facts and news (1) Status now The TA2 strand was cancelled on 28 November 2025; the programme continues. Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked.		aria.org.uk/opportunity-spaces/mathematics-for-s · aria.org.uk/opportunity-spaces/trust-everything- · aria.org.uk/insights-1/ai-progress-and-a-safegua · aria.org.uk/programme-safeguarded-ai/	verified
Centre for AI Measurement (National Physical Laboratory)Networks, bodies, events and media · also on Assurance	New centre at the National Physical Laboratory for measuring AI, launched under the government's digital and technologies sector plan. Profile Owner or parent National Physical Laboratory (NPL) Founded or opened announced 29 Jan 2026; DSIT's 10 Jun 2026 update says it was "launched"; an official launch event is listed for 10 November (year not stated) at NPL, Teddington Lab director or lead Not stated (no centre head named). Dr Peter Thompson — CEO, NPL Other leaders Chris Barnes — Head of Science for AI, NPL; Ana Alania — AI Programme Manager, NPL; James Adams and Paul Duncan — Principal Scientists, NPL (launch speakers) Site and address as published launch venue IQST, National Physical Laboratory, Hampton Road, Teddington, TW11 0LW; the centre's own location is not stated What it does An NPL-led centre that applies measurement science to how AI systems are tested, evaluated, validated, verified and monitored. The aim is to make AI assessments more robust, repeatable and comparable. It supports industry, start-ups and researchers to build and pilot technical assurance tools, and is described as a cornerstone of DSIT's AI Assurance Innovation Fund. Research groups or focus three core areas: the measurement science behind AI testing and evaluation; helping industry develop and apply assurance in real settings; coordination and capability across the wider ecosystem Funding or budget as published £10.5m public funding — 29 Jan 2026 — https://www.uktech.news/ai/national-physical-laboratory-to-establish-10m-ai-measurement-centre-20260129 (trade; NPL's own page gives no amount; single source) Partners named AI Standards Hub (NPL is a founding partner) — 2026. The launch event panel includes the Alan Turing Institute, AIQI, the Manufacturing Technology Centre, High Value Manufacturing Catapult, the AI Security Institute, Advai, Frontier Economics, the Tony Blair Institute and techUK; they are speakers, not stated partners. Uk government links DSIT AI Assurance Innovation Fund; AI Minister Kanishka Narayan quoted (Jan 2026); DSIT sector plan update (Jun 2026) Status now DSIT reported the centre launched (June 2026); NPL lists an official launch event on 10 November. Labs Round 8 deep profile (31 Oct 2026), as the researchers wrote it; confidence: verified (NPL, GOV.UK and UKTN); amount one source. Sources npl.co.uk/news/npl-to-establish-new-centre-for-a uktech.news/ai/national-physical-laboratory-to-e npl.co.uk/events/centre-for-ai-measurement-offic gov.uk/government/publications/digital-and-techn		gov.uk/government/publications/digital-and-techn · uktech.news/ai/national-physical-laboratory-to-e · npl.co.uk/news/npl-to-establish-new-centre-for-a · npl.co.uk/events/centre-for-ai-measurement-offic	one source
Centre for the Governance of AI (GovAI UK)Networks, bodies, events and media · also on Assurance	A non-profit that researches how governments and companies should govern advanced AI. Profile Owner or parent GovAI, a US 501(c)(3) non-profit (EIN 99-4000294); GovAI UK is its subsidiary, a company limited by guarantee in England and Wales (15883729) Founded or opened 2016 at Yale University; an academic centre at the University of Oxford from 2018; relaunched as an independent non-profit on 29 Oct 2021 Lab director or lead Ben Garfinkel — Executive Director (Director since 2021; Acting Director at the Oct 2021 relaunch) Other leaders John Halstead — Director of Research; Valerie Belu — Director of Talent; Katherine Gilmore — Head of Operations; Georg Arndt — Chief of Staff; Robert Trager — International Governance Lead; Markus Anderljung — Principal Researcher; Allan Dafoe — Founding Director (now listed as an affiliate) Site and address as published offices in London and Washington, DC; staff in San Francisco and remote Headcount 58 people named in the Team section (affiliates and board excluded; count made from the page) — read 31 Oct 2026 — https://www.governance.ai/people (single source) What it does A non-profit research organisation on how AI should be governed. It began at Yale in 2016, moved to Oxford in 2018, has been independent since 2021, and now has offices in London and Washington. It researches threat modelling, risk management, the economics of AI, frontier AI regulation, geopolitics, military use and technical governance. Its fellowship takes over 100 people a year. Research groups or focus Threat modeling; Risk management; Economics of AI; Frontier AI regulation; Geopolitics and military applications of AI; Military and civilian adoption of AI; Technical governance; Frontier research Notable outputs input to UK AI Security Institute research programmes, to the International AI Safety Report and to frontier companies' safety and security frameworks; publications in Science and NeurIPS (titles not given on the page) Selected papers Not stated (research page would not open) Funding or budget as published Not stated. Funders named at the 2021 relaunch: Open Philanthropy, the Centre for Emerging Risk Research and Effective Altruism Funds (no amounts) Uk government links "Input to ... UK AI Security Institute research programs" (About page) Status now Active, with offices in London and Washington, DC. Labs Round 8 deep profile (31 Oct 2026), as the researchers wrote it; confidence: one source (GovAI's own pages). Sources governance.ai/about-us governance.ai/post/the-centre-for-the-governance governance.ai/people	London (registered office EC4M 7AN)	Companies House · find-and-update.company-information.service.gov. · governance.ai/about-us · governance.ai/post/the-centre-for-the-governance · governance.ai/people	one source
ConjectureCompanies · also on Safety	London AI research company building large language models and a new AI design meant to keep advanced AI controllable and auditable. Profile Founded March 2022 Founders Connor Leahy — co-founder; Sid Black — co-founder; Gabriel Alfour — co-founder (titles not given on the fetched pages) Hq London (England: London); other_offices: null Headcount 20 — 15 February 2024 — https://www.conjecture.dev/research/conjecture-2-years What it does A London AI safety company that grew out of the open-source group EleutherAI. It started as an alignment research start-up and now describes its work as building large language models and "Cognitive Emulation" systems, meant to be bounded, controllable and auditable. Research focus AI control; Cognitive Emulation (CoEm); alignment; interpretability (e.g. "Interpreting Neural Networks through the Polytope Lens") Models and products LLM suite and Cognitive Emulation approach — as stated on About page, date null Customers or partners named null ("first enterprise customers" mentioned, unnamed, February 2024) Papers and awards research posts on conjecture.dev (e.g. "Cognitive Emulation: A Naive AI Safety Proposal") — dates not checked Rounds found 2022-04-08 (announcement) — seed-stage (stage not stated) — amount not stated — lead not stated — Nat Friedman, Daniel Gross, Patrick and John Collison, Arthur Breitman, Andrej Karpathy, Sam Bankman-Fried "among others" (NEW: DEAL-LAB-9014) Status now Website live with a London job listing; no 2025–26 news was found in this round. Labs Round 7 deep profile (30 Oct 2026), as the researchers wrote it; confidence: single_source (all company pages; no 2025–26 news found). Sources conjecture.dev/about conjecture.dev/research/we-are-conjecture-a-new- conjecture.dev/research/conjecture-2-years Selected papers (1) Interpreting Neural Networks through the Polytope Lens arXiv preprint · 22 Nov 2022one sourceProposed reading a neural network's inner workings through the regions ('polytopes') its activations fall into, which helped identify single-meaning regions.	Borough High Street, London (SE1 1JA)	Companies House · conjecture.dev/about · find-and-update.company-information.service.gov. · find-and-update.company-information.service.gov. · conjecture.dev/research/we-are-conjecture-a-new-	verified
ConsciumCompanies	London company that tests and verifies AI agents before deployment and runs a long-term research programme on machine consciousness.Key facts Products: Verifyax (live agent-verification platform), Neuroax (brain-inspired models, "Coming Soon"); research programme "Frontier" on machine consciousness; publications: Principles for Responsible AI Consciousness Research (Butlin and Lappas, 2024) and the book Perspectives on Machine Consciousness (CRC Press, ed. Chace and Lappas); open letter with 173 signatories; partnered with WPP on AI verification (release datelined "London, 18 June"; year not printed in the dateline) Profile Founders Daniel Hulme — founder, CEO and co-founder; Ted Lappas — co-founder; Panos Repoussis — co-founder; Ed Charvet — co-founder; Calum Chace — co-founder Chief executive Daniel Hulme Other leaders Not stated (only the five co-founders are named) What it does Conscium sells a platform that tests and verifies AI agents before and after they are deployed (Verifyax). It is also developing brain-inspired models (Neuroax) and runs a long-term research programme on machine consciousness (Frontier). Research focus "Neuroplastic Inference" models; machine consciousness through its Frontier programme, a "long-horizon research programme into machine consciousness". Models and products Verifyax — AI agent verification platform (simulation testing, auditing, governance), live with pricing and free trial — date not stated; Neuroax — brain-inspired models, listed "Coming Soon" — date not stated; Frontier — research programme on machine consciousness — date not stated Customers or partners named Patrick Butlin (University of Oxford) — co-drafted five principles for responsible AI consciousness research — 2024 — https://conscium.com/frontier Selected papers Principles for Responsible AI Consciousness Research (Butlin and Lappas) — PDF on conscium.com, venue not stated — 2024 — https://conscium.com/hubfs/Frontier/Principles-for-Conscious-AI.pdf; Perspectives on Machine Consciousness (book, eds Calum Chace and Ted Lappas) — CRC Press — date not stated — https://conscium.com/frontier Status now Trading; Verifyax is offered with pricing and a free trial on its website (read 31 October 2026). Note the About page gives founders' other roles — Hulme "Founded Satalia in 2012, sold to WPP in 2021 — now serving as WPP Chief AI Officer"; Repoussis "Leads WPP's AI Research Labs". No Conscium–WPP partnership is stated. Labs Round 8 deep profile (31 Oct 2026), as the researchers wrote it; confidence: one source. Sources conscium.com/ conscium.com/about-us conscium.com/frontier	London	conscium.com/frontier · fieldhouseassociates.com/conscium-launches-groun · conscium.com/ · conscium.com/about-us	one source
FACTS Grounding and FACTS Benchmark SuiteChips, models and products	Google DeepMind benchmarks that test how factually accurate language models are. released (benchmarks)Key facts Theme: AI safety research. FACTS Grounding December 2024; FACTS Benchmark Suite 9 Dec 2025.Made by Google DeepMind · 2024-12	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/7/ · deepmind.google/blog/rss.xml	lab's own pages
Gemma ScopeChips, models and products	Open tools that let safety researchers look inside Gemma 2 models to see how they process information. released (open sparse autoencoders)Key facts Theme: AI safety research. Released 31 Jul 2024. Figures: "more than 400 sparse autoencoders with more than 30 million learned features in total"; downloads on Hugging Face. Successor: Gemma Scope 2 covering the Gemma 3 family (16 Dec 2025).Made by Google DeepMind · 2024-07-31	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/gemma-scope-helping-the-saf · deepmind.google/blog/rss.xml	lab's own pages
Gemma Scope 2Chips, models and products	Open interpretability tools that let researchers inspect the inner workings of every Gemma 3 model. released (open tools)Key facts Theme: AI safety research. Released 16 Dec 2025. Predecessor: Gemma Scope (31 Jul 2024, Gemma 2).Made by Google DeepMind · 2025-12-16	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml · deepmind.google/models/gemma/	lab's own pages
London Initiative for Safe AI (LISA)Networks, bodies, events and media · also on Safety	A shared research centre in London that houses AI safety researchers and small AI safety organisations. Profile Owner or parent Safe AI London Ltd, a registered charity (1211693) and company (14942848) in England and Wales Founded or opened Not stated on pages opened Lab director or lead Krzysztof Bar — CEO — date not stated Other leaders Emma Hall — Director of Ecosystem; David Arnold — Director of Transformation; Nina Wolff-Ingham — Operations Lead; trustees Charles Ciumei, Andrew Sutton, Adam Gleave, Ryan Kidd, David Norman, Connor Campbell Site and address as published central London, "the heart of London's AI corridor" (no street address) What it does A charity that runs an AI safety centre in central London. It houses AI safety organisations, training programmes, independent researchers and visiting experts, with the stated mission of building the ecosystem needed to navigate transformative AI safely. Named residents: Apollo Research, BlueDot and IAPS. Programmes named: LASR Labs, Pivotal Research, ARENA, the Princint residency and the Tarbell Center. Research groups or focus hosted organisations and programmes as listed above Funding or budget as published backers not named Partners named Apollo Research, BlueDot, IAPS — resident organisations; LASR Labs, Pivotal Research, ARENA, Princint, Tarbell Center — programmes (dates not stated) Status now Active, per its website read on 31 Oct 2026. Labs Round 8 deep profile (31 Oct 2026), as the researchers wrote it; confidence: one source. Sources safeai.org.uk/ safeai.org.uk/team	London	forum.effectivealtruism.org/posts/4pnr37kpDKTDvu · safeai.org.uk/team · safeai.org.uk/ · lesswrong.com/posts/P7dbHRYfwykEJiDYX/announcing	one source
ShieldGemmaChips, models and products	Open Gemma 2 models that check text going into and out of AI systems for harmful content. released (open weights)Key facts Initial release 31 Jul 2024; sizes 2B, 9B and 27B; built on Gemma 2; targets four harm types (sexually explicit, dangerous content, hate, harassment); paper arXiv 2407.21772; access: open weightsMade by Google DeepMind · 2024-07-31Versions Gemma → Gemma 2 (and CodeGemma, RecurrentGemma, PaliGemma → PaliGemma 2, ShieldGemma → ShieldGemma 2, DataGemma)	London	ai.google.dev/gemma/docs/shieldgemma/model_card · ai.google.dev/gemma/docs/releases	lab's own pages
ShieldGemma 2Chips, models and products	Open model that checks images for harmful content against set safety policies. released (open weights)Key facts Released 10 Mar 2025; 4B-parameter image safety model built on Gemma 3; access: open weightsMade by Google DeepMind · 2025-03-10Versions Gemma → Gemma 2 (and CodeGemma, RecurrentGemma, PaliGemma → PaliGemma 2, ShieldGemma → ShieldGemma 2, DataGemma)	London	ai.google.dev/gemma/docs/releases · deepmind.google/models/gemma/shieldgemma-2/	lab's own pages
Sparrow (dialogue agent)Chips, models and products	Research chatbot trained with human feedback and rules to give evidence-backed answers and avoid unsafe replies. research (paper only)Key facts Announced 22 Sep 2022; gave a plausible answer supported by evidence 78% of the time on factual questions; participants tricked it into breaking rules 8% of the time; paper arXiv 2209.14375; access: paper onlyMade by Google DeepMind · 2022-09-22	London	deepmind.google/blog/building-safer-dialogue-age	one source
SynthID DetectorChips, models and products	A portal to help identify AI-generated content marked with SynthID. released (portal)Key facts Theme: AI safety research. Announced May 2025 (index). Related: AI image verification in the Gemini app (November 2025); content-provenance post (17 May 2026).Made by Google DeepMind · 2025-05	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/7/	one source