Hey folks! Apologies for the delay in getting this post out: I was in Geneva with the Emerging Leaders in Biosecurity Initiative (great programme, highly recommended and thanks very much to the Johns Hopkins Center for Health Security for running it!) but pleased to be able to share more specific thoughts on def/acc with you all.
TL;DR: The UK needs to privilege defensive actors with early access to frontier AI capabilities across biological defence, intelligence and AI security. I outline 19 specific interventions—including Trusted Tester programmes, AI-enabled microbial forensics, personalised red-teaming for intelligence analysts, and automated software security. AI agents are doubling their task-solving horizon every four to seven months, so the scale of the risks and the opportunities is growing very, very quickly. Securing AI means getting projects off the ground now and executing in 2026. Governments must allocate compute and establish privileged access pathways for defensive actors. Industry needs to develop secure deployment infrastructure to get them frontier AI. Funders should prioritise measurable and deployable defensive capabilities. Let’s get building.
“UK Sovereign AI should lead the delivery of a government offer to new and existing frontier AI companies that includes … [f]acilitating deep collaboration with the national security community”
– AI Opportunities Action Plan, Matt Clifford, 13 January 2025
Accelerating strategic defensive capabilities with AI is robustly good. There are clear economic benefits—new medicines, new start-ups (and fewer billions-scale post-cyber-incident loans …)—as well as opportunities to double-down on UK strategic advantage in cybersecurity, intelligence gathering and analysis, computational biology and AI security and assurance. The UK Government recognises this: the mission of its Sovereign AI Unit is to “harness AI’s capabilities to unlock economic growth and enhance UK national security”.
… AI-enabled def/acc is also entirely necessary in a world full of adversaries who will increasingly adopt AI to try to realise their own security ambitions.
Here, I suggest 19 concrete defensive acceleration projects that I would be excited to see the UK further develop. These are organised into three themes:
chemical and biological defence
intelligence advantage
AI security
Underpinning each of these is a “zeroth” theme, that of privileging defenders, which is perhaps the most important. Although I am confident that these ideas are beneficial for the UK’s economy and national security, it is hard to pick winners from the centre of government. This means that enabling public and private defenders to use frontier AI—facilitating that deep collaboration, in Clifford’s words—is the crucial first step.
0 Foundations: how to privilege defenders
The most straightforward way to tip the offence–defence balance in the right direction is to give defenders better and earlier access to frontier AI systems before public release. Everything else follows from this: no access equals no defensive uplift.
This means deliberately privileging specific actors with prototype capabilities. This includes providing de-safeguarded or ‘maximally helpful’ AI models: safety refusals will occasionally need to be turned off to enable legitimate dual-use research. For chem/bio defence, the groups I have in mind include technical experts, like countermeasure developers and detection and microbial forensics experts, but also cybersecurity teams and intelligence analysts.
Google DeepMind pioneered this approach with their AlphaProteo and AI co-scientist Trusted Tester programmes, sharing their advanced AI–bio capabilities with world-class academic labs to validate the models and help improve them.
Trusted Tester programmes are good for def/acc because they solve two problems simultaneously.
Defensive actors gain months of headstart over potential adversaries, learning to integrate new capabilities into their workflows while threat actors have to wait for the public release. A vaccine developer who has spent six months experimenting with an advanced protein design tool will achieve far greater uplift than someone accessing it for the first time on release day.
AI developers receive crucial feedback from expert users working on real defensive problems—information they cannot get from internal testing alone. This feedback loop accelerates both the refinement of the tool and the defenders’ ability to use it effectively.
The UK should systematically expand this model across national security priorities. This requires three components:
Vetting infrastructure: Government must coordinate a trusted network of defensive actors who should receive privileged access. Civil servants will likely already be vetted to the required level, but academic and industry defenders may need additional assurance. In some cases, something like the government’s Baseline Personnel Security Standard (or equivalent background checks) will do, but there will be actors who require more advanced capabilities and need to be cleared properly (e.g. to Counter Terrorist Check, the first of the three National Security Vetting levels). (The most sensitive research will require much higher levels of clearance, but this can be managed in the usual way by the Government, internally.)
The Cabinet Office (CO) and the Department for Science, Innovation and Technology (DSIT) should work together to initiate and prioritise these clearances.
DSIT can sponsor the clearances themselves, but CO should formally support these applications on the grounds of addressing pressing national security priorities and should work to ensure world-class experts are cleared quickly to access frontier AI capabilities.
Secure deployment systems: Many defensive applications require capabilities that would be dangerous if made publicly available. This means building secure API infrastructure that enables sophisticated use without releasing model weights. For defenders out of government, this likely involves (eventually) assuring systems to RAND Security Level 3 (to thwart well-resourced terrorists and insider threats). Government defenders will need—and often, should already have—higher levels of security.
The National Cyber Security Centre and AI Security Institute should work with frontier AI companies to ensure this infrastructure meets these standards.
Measurement frameworks: We should actually verify that Trusted Tester programmes generate defensive uplift. This looks like running randomised controlled trials (RCTs) comparing defensive actors with and without early access, measuring time-to-capability, quality of outputs and integration into operational workflows. Without measurement, we’re just hoping that giving people tools makes them better at their jobs.
I recommend that the Government Digital Service extend their previous RCTs since they have a proven track record of delivering in this space.
Eventually, these trials should also be run on the government’s classified systems—the ‘high-side’—but the significant complexity of this environment makes it sensible to pilot on unclassified networks first.
That said, Trusted Tester programmes must extend beyond academic researchers to those working on classified national security projects. Current programmes allow biologists to email alphaproteo@google.com to express interest in testing protein design capabilities. We need equivalent pathways for Defence Science and Technology Laboratory (Dstl) scientists designing countermeasures for engineered pathogens, and intelligence analysts building threat detection systems and conducting operational planning. These users cannot publicly announce their participation or describe their use cases, which likely requires different vetting mechanisms and security arrangements—but their defensive uplift matters just as much as academic research, if not more.
Other privileging interventions
Beyond Trusted Tester programmes, the UK could consider additional complementary approaches to privilege defenders:
Defensive compute allocation: Reserve sovereign compute capacity for internal government defensive projects beyond allocations to defence technology companies. This ensures that Dstl, UKHSA, and especially the Intelligence Community and other national security organisations can run compute-intensive defensive research without competing for commercial cloud resources. It is currently unclear whether the AI Research Resource (AIRR), which is available to ‘public researchers’, is sufficiently secure for this purpose. It is also important to reserve compute for R&D and not just routine defence and security activities, so I expect that existing compute allocations for the UK Intelligence Community need to increase.
Defensive embargos: Restrict public release of advanced AI capabilities until defensive actors have assessed offensive and defensive applications and incorporated defensive applications into their workflows. This would create a deliberate 6–12 month capability advantage for defence and security organisations. I appreciate this is a bigger ask but getting the systems and authorisation in place now is really important, in case we want to enact this tiered privilege scheme in the future—see this paper from colleagues at CLTR for more discussion of AI-related emergency powers.
Government adoption of best models: Ensure UK departments use frontier AI capabilities as standard, including the best open-weight models after security vetting. Update rapidly as new capabilities emerge rather than remaining locked into outdated systems due to procurement delays. (And maybe stop using government funding for research into whether a Llama 3.3 is better than Llama 3.1 or if o1 is smarter than GPT-4o—here are three such examples—given that these just confirm what public benchmarks already tell us: that the latest models are marginally better than the last. Learning by doing is more important than chasing small efficiency gains.)
1 Chemical and biological defence
The biological domain is fundamentally offence-dominant—pathogens self-replicate and spread whilst countermeasures must be manufactured at scale, attacks are cheap whilst defence is expensive, and adversaries can iterate until success whilst defenders cannot afford even one failure. This means that defence-neutral AI tools, no matter how scientifically impressive, will not improve our security position in absolute terms. Risks are increasing, so we must build and deploy defence-dominant capabilities quickly.
One easy win for the UK is to accelerate the UK Microbial Forensics Consortium (UKMFC) with frontier AI capabilities for sequence analysis and attribution.
The UKMFC is already a world-leading capability and fast, reliable attribution is the foundation of effective deterrence in the biological domain, as well as enabling proportionate response to biological attack.
Accelerating the UKMFC means deploying AI systems that can rapidly analyse microbial samples, flag engineered signatures and identify the lab-of-origin for that sample. There is already early proof-of-concept work for plasmid sequence lab-of-origin investigation, which we should extend to viruses, taking advantage of huge leaps in protein and DNA language model design. There was even a genetic attribution challenge a few years ago before the rise of language models. I expect that whoever tries this now will smash the state-of-the-art, so the UK should grab this low-hanging fruit before someone else does.
Who should act:
Government: Dstl should lead technical development for UKMFC, working with AISI to provide privileged access to frontier AI models. They should also follow-up their first Defence and Security Accelerator (DASA) call on Sample Preservation, Analysis and Attribution with a second one focused explicitly on AI-enabled methods that automate signature detection for engineered pathogens (including ones made with AI-generated designs).
Industry: UK companies developing AI sequence analysis tools should work with UKMFC as early customers, incorporating requirements for forensic-grade outputs, where appropriate. These technologies also likely have commercial applications beyond defence—such as in pharmaceutical quality control, agricultural pathogen tracing and environmental contamination investigation.
Funders: The second DASA call should fund specific AI acceleration projects for UKMFC capabilities. While UKMFC does need to be able to produce the highest-validity forensic evidence to satisfy a court, I recommend first trying to reduce the time until the Prime Minister is confident they can stand up in Parliament to attribute an attack, like Theresa May did after the Salisbury novichok attacks. Yes, eventually we’ll require a higher level of assurance, but the strategic advantage the UK needs now is providing the most senior decision-makers with accurate information, as quickly as possible. It’s likely that philanthropic funders could also fund a lot of the basic (unclassified) research outside of the Ministry of Defence, too.
Other counter-CB defence options:
Anti-hospitalisation pandemic virus vaccine design: Build end-to-end AI-enabled pipelines for vaccine and therapeutic design that generate optimised candidates in minutes rather than months. I recommend piloting the development of “design” stockpiles for pandemic viruses. These should be candidate multi-valent vaccines (that protect against many different strains for a given pathogen, like SARS-1, SARS-2 and MERS) that have been shown to be safe in Phase 1 trials and are predicted to prevent severe disease. The UK is already one of the top countries in the world for some of this work, with recent papers on influenza, on mpox and several on coronaviruses. The most important outcome to prevent is a deliberate attack that knocks out the ability for the UK Armed Forces and other essential workers to respond effectively: we need to be able to confidently tell adversaries “No matter what you throw at us, the UK is resilient and we’ll be able to respond in force” because society can continue to function, even if most citizens are staying at home to avoid disease.
Broader 100 Days Mission acceleration: Eventually, you need to do the above for all pathogens and across all parts of the 100 Days Mission (that aims to deliver vaccines, diagnostics and therapeutics within 100 days after the detection of a new pandemic). This is so important, but will take a lot of time and we’ll have to learn by doing, so I’d start with the pandemic virus pilot above.
AI-enabled environmental metagenomic surveillance: Build the UK version of the Nucleic Acid Observatory: pathogen-agnostic surveillance through metagenomic sequencing of environmental samples (wastewater and air) combined with AI anomaly detection algorithms. The UK can leapfrog the U.S. by starting with an AI-first approach, using AI systems trained on historical sequencing data which rapidly analyse billions of daily reads, flagging suspicious sequences that diverge from baseline patterns (including those from deliberately engineered pathogens designed to evade conventional screening). Advanced protein-folding models can then assess whether unknown sequences encode dangerous functional domains. The UK already has strong metagenomic capabilities through UKHSA’s mSCAPE clinical surveillance programme; extending these to systematic environmental sampling at airports, transport hubs and urban centres would provide early warning for biological threats before they cause widespread infection.
Retooling existing capabilities: UKHSA currently uses AI-enabled bioaerosol analysis for … pollen monitoring. I hate hayfever as much as the next person, but I dislike pandemics even more. We should redirect this and similar capabilities towards higher-priority threats where the strategic value might be just a smidge bigger.
Automated narrow AI–bio tool evaluation: Build systems that automatically classify proposed AI–bio tools as defence- or offence-dominant before funding, building and publishing them. UKRI has just released new guidance on dual-use research, so now is the time. It can be really difficult for academics—who don’t have security training—to work out whether proposed AI capabilities are strategically beneficial for UK defence. Rather than just telling them to figure it out anyway, we should be building state-of-the-art classifiers that automate decision-making and free up academic time to focus on developing exciting new capabilities. RAND Europe and CLTR already piloted such a system in our recent Global Risk Index, so the UK is primed to build the infrastructure here to serve the global AI–bio tool economy.
Nucleic acid screening automation: Automate the sequence and customer screening components of nucleic acid screening to make it cheaper and easier for biotech companies to serve legitimate customers but reject malicious orders from criminals and terrorists. Good for defence and good for security (and we have really strong protein language models now, this is totally doable!). The UK already has some of the strongest screening guidance in the world and I personally know multiple PhDs who are interested in founding a UK start-up to do this. A tiny bit of seed funding plus some security expertise (that only the Government can provide) could effectively guarantee that the UK captures a slice of the £4bn global synthesis market.
2 Intelligence advantage
(In this section, I’m suggesting capabilities that I am confident that the UK, and many other countries, will need very soon. It may be that the UK already is doing this work, in which case, fantastic! I’d be keen to discuss ongoing efforts to enhance UK intelligence collection, analysis and dissemination efforts with AI, but appreciate that this gets very sensitive, very quickly. Apologies in advance if UKIC colleagues read this and go “Yep, done that … and that … and that”, but I fear that this is not wholly the case.)
The UK Intelligence Community produces analysis and assessment to help inform Ministers on consequential national security decisions. These products must withstand intense scrutiny from policymakers, allies and (it is often correct to assume) hostile actors seeking to exploit analytical gaps. Receiving feedback is understandably tricky for analysts: information is siloed according to need-to-know and the work can get very technical, very quickly with only a small number of analysts working in some, less prominent areas. AI can potentially strengthen analytical rigour by providing adversarial challenge at scale via personalised red-teaming for analysts.
I propose that the UK deploy secure AI instances that roleplay as colleagues, senior leaders or intelligence customers to critique intelligence products before dissemination. Analysts submit draft assessments to AI systems trained to identify cognitive biases, challenge assumptions, probe gaps in reasoning and suggest alternative explanations. The AI could adapt its challenge style based on the user’s role—mimicking peer review for junior analysts, executive-level questioning for senior scrutiny, or other customer requirements for tailored products.
Something like this capability is eventually going to exist in every business in the country. Forward-looking companies are already encouraging their employees to use LLMs liberally to quality-check and improve their work. But getting a system that is sufficiently secure and optimised for the very specific activities that take place only in the Intelligence Community is going to be much harder. So we need to start the work now.
There’s a nice economic argument here: fine-tuned, secure models that work for UKIC can easily be repurposed for other business activities (and sold to close allies and partners), but it probably doesn’t work the other way around. And strengthening and automating analytical rigour is no bad thing for the UK’s standing with Five Eyes partners: yes, everyone agrees the UK has world-leading intelligence capabilities, but being the first adopter of AI at scale could potentially help the UK be the undisputed leader. Making sure the UK remains at the front of the AI-enabled intelligence race is also crucial to ensure the UK maintains a strong voice in future U.S.-led AI–national-security decisions in the next few years.
Other intelligence interventions
Automating integration in all-source threat assessments: Deploy AI agents to combine and analyse material used in all-source briefs. Begin with open-source and unclassified material, then extend to the high-side.
Expanded early warning automated assessment: Extend early warning assessment using AI to generate proactive threat analysis. PHIA guidance does allow for intelligence assessment bodies to produce assessments without explicit tasking when early warning value is high—AI agents seem like the obvious way to deliver this at scale. The National Situation Centre already integrates real-time open-source information, supported by the National Security Watchkeepers, but there’s reason to think AI could really be superhuman in this domain: AI systems can potentially absorb all news and social media in the world simultaneously (not just look at a few feeds on a big wall of screens) and integrate them in seconds. Again, that sort of capability will have much wider applications across the economy—it’s basically just “the ability to integrate global information quickly” … quant hedge funds make billions from doing this!—but getting it right in the defence domain first is the important thing.
Privacy-preserving cross-domain integration: Build AI query systems that enable analysts to search across compartmented material without revealing unnecessary information between compartments. This reduces barriers to all-source integration whilst maintaining security protocols. (Yep, this is definitely a more speculative capability, but I was talking to someone recently about it and advances in federated learning suggest it might actually be possible soon. The UK should be willing to take risks on these sorts of blue-sky ideas.)
3 AI security
As AI systems become more capable and widely deployed, securing them against misuse, theft and manipulation is already becoming a critical national security priority. This includes both defending AI systems themselves and using AI to defend other critical infrastructure.
Automated software security at scale
We should deploy AI systems to automatically refactor legacy code for memory safety and patch known vulnerabilities in open-source software. Memory-safety vulnerabilities—buffer overflows and use-after-free errors—account for roughly 70% of critical security bugs in widely-used software. Current approaches require substantial manual effort to refactor millions of lines of code in critical infrastructure. AI agents can automate much of this work, starting with high-priority systems and expanding coverage systematically. I found exactly one existing effort to do this: The Great Refactor led by Herbie Bradley, a UK PhD student (who blogs at AI Pathways) and used to work at the UK AI Security Institute. This effort is supported by a U.S. think tank, Institute for Progress. I expect there is room for more organisations trying to do this, and support from UK VCs and the UK Government. I’m also very excited about existing efforts like CHERI—a joint venture between University of Cambridge and SRI in Silicon Valley, which is supported by DSIT—but the latter works at the hardware level, while refactoring legacy code is a software-level intervention. We’ll likely need both approaches.
Fixing memory-safety vulnerabilities really matters—the UK and U.S. governments are both very clear on this—because our critical national infrastructure runs on software full of them. Adversaries continuously scan for and exploit these weaknesses—both the WannaCry attack in 2017 and the CrowdStrike outage in 2024 were consequences of poor memory safety. We cannot manually secure all critical code before AI agents make exploitation easier, so we must use AI agents for defence first. This creates a genuine race: will defensive or offensive actors deploy capable coding agents at scale faster? Getting memory-safe refactoring deployed across government systems and critical infrastructure within the next 18 months could prevent substantial harm.
Who should act:
Government: A DASA call here could work very well, coordinated between DSIT and MOD with technical advice from the NCSC. Eventually, NCSC should lead deployment across government systems, starting with highest-risk infrastructure and then coordinating with critical national infrastructure operators to extend coverage beyond government.
Industry: UK companies developing AI coding assistants should prioritise security refactoring capabilities. This has both defensive value and substantial commercial potential: every major software company faces the same technical debt problem.
Funders: The Great Refactor project estimates it could cost $100mn, which means potentially a single cybersecurity-focused billionaire could actually get this done, so large philanthropies and VCs should also be paying attention here. Forbes estimates that the global cybersecurity market is around $200bn (and growing rapidly) …
Other AI security interventions
AI-enabled tabletop exercises: Build AI systems that generate realistic threat scenarios for government exercises. These systems can create more challenging, varied and unpredictable exercises than human scenario developers and should eventually be orders of magnitude cheaper and faster. Given the UK’s ambition to deliver a Tier 1 National Exercise every year out to 2030, automating smaller-scale exercises seems sensible. (On a personal note, I’d love to see this capability just for the inevitable reapplication to Dungeons and Dragons …)
Automated real-time analysis of open-weight models: Deploy systems that continuously monitor newly-released open-weight models for dangerous capabilities. This enables rapid warning when concerning models appear and informs export control decisions. AISI should coordinate this capability for the UK government (and I expect there’s already a kernel of a capability here, but much more that the private sector could do to build the right scaffolding and infrastructure).
Sleeper agent detection: Thanks to really excellent work from the AI Security Institute and close partners, we now know that data poisoning is actually much easier than we thought (eek!) We’re going to need automated data poisoning and sleeper agent detection systems. This is an economy-scale opportunity because, eventually, we will deploy AI agents everywhere that there’s a computer. So the commercial applications are enormous, but getting the cybersecurity right likely needs government support. A DASA call, joint with NCSC, seems the obvious next step.
Conclusion
Defensive acceleration requires building defence-dominant capabilities, privileging defensive actors with early access, and measuring whether we achieve genuine uplift. The interventions outlined here meet these criteria—but only if we execute rather than deliberate further.
Trusted Tester programmes must come first. Without systematic early access to frontier AI, defensive actors cannot achieve the uplift that everything else depends on.
There are exciting and impactful interventions across all three priority areas above. First, accelerate the UK Microbial Forensics Consortium with AI-enabled attribution—an easy win building on world-leading capability. Following this up with AI-enabled environmental metagenomic surveillance at transport hubs and urban centres strengthens early warning before widespread infection and positions the UK ahead of U.S. efforts. Second, start figuring out how to rapidly accelerate aspects of intelligence analysis and assessment, especially through local, secure red-teaming. Third, automate software security at scale—adversaries will deploy coding agents to exploit vulnerabilities, so we must deploy them for defence first.
Of course, making this a reality will mean overcoming significant implementation hurdles, from well-known public sector procurement bottlenecks to skills gaps and cultural resistance to adopting new technology at pace. But these are surmountable with effort commensurate with the scale of the risks and the opportunities.
We face a closing window. AI agents double the length of the task they can solve every four to seven months. Whether AI favours defence or offence depends on choices made in 2025 and 2026. The UK can maintain strategic advantage by moving first—but this requires execution, and fast.
Let’s get building.
If you made it to the end of this bumper edition, thank you so much! And thank you especially to Cassidy Nelson and Eleanor Hevey for the helpful review. This brings me to the close of my Asterisk AI Fellowship, but please do comment below or message me with suggestions on future posts and thanks for reading!


