governments and the AI industry<\/a> need to be better prepared to stop a nightmare scenario at some point in the future.<\/p>\nPopular large language models (LLMs), such as ChatGPT, \u201ccould soon drastically lower the informational barriers for planning and executing biological attacks\u201d, by inadvertently \u201chelping novices develop and acquire bioweapons by providing critical information and step-by-step guidance\u201d, a report by the authoritative Center for Strategic and International Studies (CSIS) warned earlier this year.<\/p>\n
OpenAI, which owns ChatGPT, and other tech firms have stepped up their safeguards in response to the threat, the CSIS report found. The i Paper has approached the tech firms for further comment.<\/p>\n
Assessments from leading AI labs \u201cdemonstrate that LLMs are rapidly approaching or even exceeding critical security thresholds for providing users key bioweapons development information \u2013 with some models already demonstrating capabilities that surpass expert-level knowledge\u201d, the report said.<\/p>\n
It added: \u201cWhile some of the leading companies are voluntarily imposing safeguards, the overall trajectory nevertheless points toward a near-term future in which policymakers must confront bioterrorism risks not just from sophisticated state and terrorist organisations, but potentially from individuals with little technical background but access to popular LLMs.\u201d<\/p>\n
\u2018Pandemic-scale pathogens\u2019<\/p>\n
Future AI biological design tools, or BDTs, could \u201cassist actors in developing more harmful or even novel epidemic or pandemic-scale pathogens\u201d, which could evade existing safeguards, the report also found.<\/p>\n
Under one scenario raised by the CSIS report, an existing pathogen such as a strain of avian influenza \u2013 which has a high fatality rate in humans but cannot currently transmit between people, meaning it is less of a threat \u2013 could be adapted to spread more easily.<\/p>\n
A safety assessment published in February 2025 of Deep Research, one of OpenAI\u2019s advanced AI capabilities, which was cited in the report, found that its models \u201care on the cusp of being able to meaningfully help novices create known biological threats\u201d.<\/p>\n
OpenAI said at the time it expected \u201ccurrent trends of rapidly increasing [biological] capability to continue, and for models to cross this [high-risk] threshold in the near future\u201d \u2013 meaning that the models could potentially allow misuse.<\/p>\n
In preparation, it said, the company is \u201cintensifying our investments in safeguards\u201d. <\/p>\n
Another biological design tool cited in the CSIS report, from AI chip designer Nvidia, known as Evo 2, has been trained on 128,000 genomes from humans, animals, plants, bacteria and some viruses and allows scientists to identify novel patterns between different DNA sequences.<\/p>\n
The research model has already identified, with 90 per cent accuracy, whether different mutations of a breast cancer-related gene are benign or not.<\/p>\n
\u2018We need to think about our existence\u2019<\/p>\n
Barrons told The i Paper that great progress was being made in medicine using AI to tailor diagnosis and treatment for individuals. <\/p>\n
But he added: \u201cThe most terrifying thing is the same technology that makes individually targeted medicines could make individually or targeted pathogens. <\/p>\n
\u201cAnd one scary side of that is a state does it \u2013 so a bio-war. So essentially, this pathogen targets a characteristic of the people they don\u2019t like.<\/p>\n
\u201cAnd that might be what you can imagine \u2013 it can be race, it can be colour, whatever that\u2019s horrible.<\/p>\n
\u201cOr scary, scary AI, which drives these things for whatever reason, produces and releases a pathogen that\u2019s designed to kill all humans, and to which there is no antidote, because, in its own sweet way, the AI has worked out we control the off button.<\/p>\n
\u201cNone of that is imminently likely, but what it does say is \u2026 we need to think about the risk to our security \u2013 indeed \u2013 our existence.\u201d<\/p>\n
Covid-19 \u2018would look like the common cold\u2019<\/p>\n
Indeed, the CSIS report warned \u201cif BDTs such as Evo 2 can generate good outputs, then presumably more capable future tools could create bad ones, too\u201d.<\/p>\n
It continued: \u201cImagine, for instance, bird flu: Some strains have a reported human mortality rate of more than 50 per cent, meaning that approximately one out of every two people that contract the disease will die.<\/p>\n
\u201cThere are other versions of flu that are not as lethal but are more contagious. Someone who contracts the more common seasonal flu, for instance, will spread it to 1-2 other people on average. <\/p>\n
\u201cMalicious actors, however, might use a more advanced BDT with generative capabilities to create a new version of bird flu that is both highly lethal and highly contagious. <\/p>\n
\u201cSuch a scenario would likely make the Covid-19 pandemic \u2013 which has claimed more than 27 million lives and shaken the world economy \u2013 look like the common cold.\u201d<\/p>\n
Threat is \u2018marginal\u2019 but \u2018high ceiling of harm\u2019<\/p>\n
However, experts have urged caution over what might be possible in the near future.<\/p>\n
Caleb Withers, research associate for the Technology and National Security Programme at the Center for a New American Security, said: \u201cThere\u2019s not some distinct class of AI-enabled biological weapons that you might see \u2013 ultimately, DNA is DNA and RNA is RNA.<\/p>\n
\u201cAnd so it\u2019s more about how people get to discovering or weaponising certain agents. And the caveat is that largely most of what I\u2019m describing is in the future tense here, and I don\u2019t think there\u2019s necessarily an acute threat.\u201d<\/p>\n
AI models are \u201cincreasingly useful at helping people be more effective at various tasks\u201d, which also applies to biological science, Withers said, but this was \u201cmarginal rather than transformative at the moment\u201d.<\/p>\n
He added: \u201cThe concern here is that if you consider the task of trying to develop a bioweapon \u2026 you can better predict or model what changes to organisms might lead to these effects.<\/p>\n
\u201cThat\u2019s obviously going to potentially result in people being able to more effectively make more dangerous biological agents.\u201d<\/p>\n
He added: \u201cAt the limit, bioweapons can be pretty devastating. <\/p>\n
\u201cThe ceiling of harm is pretty high. The caveat there is that it\u2019s not that straightforward to just discover and develop such a biological agent.<\/p>\n
\u201cThe current effect of AI is pretty marginal, but AI is getting better rapidly, faster at relevant tasks.\u201d<\/p>\n
Withers added it could pose a \u201cpretty acute challenge\u201d over the next few decades.<\/p>\n
\u201cI wouldn\u2019t want to overstate the current threat. But I wouldn\u2019t want to understate how fast AI is improving or understate how much damage could be caused at the limit, through pathogens, and\u2026 the importance of preparedness for this sort of thing. <\/p>\n
\u201cAnd so I think it just behoves thoughtful monitoring by governments about how fast are the relevant capabilities actually developing\u2026 it\u2019s about having well-resourced coordinated efforts to stay safe.\u201d<\/p>\n
AI could help bad actors \u2018hide\u2019 activity<\/p>\n
Dr Alexander Ghionis, research fellow in chemical and biological security at the University of Sussex Business School, said AI does not \u201cdesign\u201d a biological weapon in a \u201csingle decisive step\u201d and would always require some form of human involvement that is labour-intensive and involves compromises.<\/p>\n
He said: \u201cThe familiar idea of AI enabling population-specific or geographically targeted biological weapons is misleading: even if such science were available \u2013 and it is not \u2013 someone would still have to solve all the labour-intensive problems in between.\u201d<\/p>\n
AI can, however, accelerate certain forms of labour such as carrying out straightforward scientific tasks like finding papers, drafting protocols, troubleshooting, as well as non-scientific tasks like planning, procurement, and logistics.<\/p>\n
It could also help actors \u201chide\u201d their activity, which could make oversight by industry or governments more challenging, Ghionis said.<\/p>\n
\u201cThe genuine effects of AI tend to sit around the edges \u2013 in how actors imagine, plan, troubleshoot, coordinate, conceal, or experiment \u2013 rather than in any radical change to what biology allows.<\/p>\n
\u201cAI can reshape parts of the labour of harm \u2013 altering speed, sequencing, and perceived feasibility \u2013 but it never removes the biological, organisational, or practical constraints that make these weapons so difficult to acquire and use, whether the intended effects are lethal or merely incapacitating.<\/p>\n
\u201cThe biological danger itself does not change because AI is involved. Biological agents remain what they are: fragile, unpredictable, and constrained by the environments in which they are expected to thrive. <\/p>\n
\u201cAI does not suddenly unlock new categories of pathogens, nor does it make biology behave with the precision that is sometimes imagined.\u201d<\/p>\n
Risks outweigh benefits for hostile actors<\/p>\n
Ghionis said that there were still many risks for anyone wanting to develop a bioweapon, and while AI might streamline certain things in the process, biological agents were still difficult to control.<\/p>\n
He added: \u201cThe moment an actor cannot reliably predict who will be affected, for how long, and under what conditions, the strategic value of an incapacitating agent quickly falls away. And nothing about AI resolves the fragility or unpredictability that undercuts biological controllability\u2026<\/p>\n
\u201cOnce you look closely at intent and utility, the number of actors for whom an uncontrollable, highly lethal biological agent makes strategic sense is very small. <\/p>\n
\u201cThe risks \u2013 blowback, loss of control, political isolation, economic self-harm \u2013 far outweigh any plausible benefit for most states, and for many non-state groups as well.<\/p>\n
\u201cEven incapacitating agents, which seem more realistic on paper, quickly run into problems of controllability and reliability when you try to imagine their use in real-world conditions.\u201d<\/p>\n
In fact, the greatest biological risks the human race faces still come from natural outbreaks rather than engineered ones, Ghionis said.<\/p>\n
\t\tYour next read<\/p>\n
\n\t\t\t![\"Article](\"https:\/\/www.newsbeep.com\/uk\/wp-content\/uploads\/2026\/01\/SEI_278289562.jpg\")
<\/a><\/p>\n\u201cAnd the measures that protect societies from natural threats \u2013 strong surveillance, resilient health systems, capable laboratories, well-trained first responders, and effective coordination across agencies \u2013 also form the best defence against deliberate misuse.<\/p>\n
\u201cA biological weapon becomes far less attractive if its expected effects can be blunted or rapidly contained.<\/p>\n
\u201cSo I would say the UK is moving in a sensible direction, but sustained investment is essential: in local and national health resilience, in scientific capacity, in international engagement, and in the steady diplomatic work that reinforces the norm against biological weapons.\u201d<\/p>\n
Sir Keir Starmer\u2019s government published a new UK Biological Security Strategy in 2025 which acknowledged that \u201crapid developments in AI-enabled scientific tools and engineering biology offer significant opportunities for drug discovery and vaccine development, but also have the potential to introduce new risks of misuse\u201d.<\/p>\n
The government has set up a review to evaluate \u201chow advanced AI could assist chemical and biological misuse, and worked with companies to address risks and develop their own thresholds\u201d.<\/p>\n","protected":false},"excerpt":{"rendered":"Large language models like ChatGPT could soon drastically lower the informational barriers for planning and executing biological attacks…\n","protected":false},"author":2,"featured_media":347272,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[554,733,4308,198,3052,86,56,54,55],"class_list":{"0":"post-347271","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-artificial-intelligence","8":"tag-ai","9":"tag-artificial-intelligence","10":"tag-artificialintelligence","11":"tag-biological-warfare","12":"tag-defence","13":"tag-technology","14":"tag-uk","15":"tag-united-kingdom","16":"tag-unitedkingdom"},"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts\/347271","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/comments?post=347271"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts\/347271\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/media\/347272"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/media?parent=347271"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/categories?post=347271"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/tags?post=347271"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}