Britain’s top spy has named the AI risk that worries him most, and it isn’t deepfakes or hacking. It’s AI bioweapons: the chance that AI could make it easier to build a biological weapon, a risk MI5 chief Sir Ken McCallum says could “change the stakes”.
He said it in his annual threat speech on 13 October 2026. A week later, the Guardian’s Science Weekly put the question to one of the world’s leading biosecurity experts. Here is what was said, why the worry is reasonable, and what is actually being done about it.
Key takeaways
- On 13 October 2026, MI5 Director General Ken McCallum said AI has the “potential to change the stakes on existing or even new bioweapons”.
- He described this as the AI misuse he weighs most heavily, while stressing that it has not happened yet.
- MI5 has opened an AI Threat Assessment Centre that pulls in experts from the AI Security Institute, the National Crime Agency and the National Cyber Security Centre.
- The main fix on the table is mandatory screening of synthetic DNA orders. AI lab chiefs backed it in June 2026, and a US bill to require it already exists.
What did MI5 say about AI bioweapons?
McCallum gave the speech at Thames House, MI5’s London headquarters. According to AFP’s report via Digital Journal, he said AI is “already supercharging the cybersecurity contest between attackers and defenders”.
He described biology as the bigger long-term stake. He also made a broader point that sticks: “human history has given us no experience of facing an actor which is smarter than we are.”
This was not abstract. ITV News reported that MI5 has already found suspects using AI to improve their reconnaissance and attack planning, including with AI-enabled drones.
Why do experts say these risks are real?
The Guardian’s Ian Sample interviewed Dr Tom Inglesby, director of the Johns Hopkins Center for Health Security. The episode’s title quotes him: “These risks are real.” Inglesby has been making that case for years.
In written testimony to a US Senate subcommittee in November 2023, he named two concerns at the overlap of AI and biology. One was AI helping to bring back dangerous viruses. The other was AI helping to design new agents that could start a pandemic. He urged lawmakers to make strong governance of both a high priority.
The core worry is what experts call the “knowledge barrier”. Making a biological weapon has always taken scarce expertise, tacit lab skills and many failed attempts. Think of it like a locked door with several bolts. AI doesn’t kick the door down. The fear is that it slides back one or two bolts by answering questions that once needed a trained specialist.
Two kinds of tools are in the frame. General chatbots, known as foundation models, can explain and troubleshoot. Specialist biological design tools can propose new proteins or sequences. Each raises a different problem for anyone trying to set guardrails.
What is MI5’s new AI Threat Assessment Centre?
The new centre sits inside MI5 and draws on the AI Security Institute (AISI), the National Crime Agency and the National Cyber Security Centre. McCallum said its job is to produce all-source assessments that “separate the bluster from the realities”.
The timing matters. Resultsense reported that on the same day, OpenAI and Anthropic told the House of Commons business committee that a warning from MI5, MI6, GCHQ or AISI would lead them to halt a model. A team inside MI5 that can provide that kind of evidence becomes a real lever.
Some details are still unknown. MI5 has not said how big the centre is, who it reports to, or whether its findings will go to AI companies directly or through AISI.
How can AI bioweapon risks be reduced?
No single measure solves this. The current approach stacks defences in layers, so a failure in one can be caught by another. Here’s how the main layers compare.
| Layer | What it does | Status in 2026 |
|---|---|---|
| Model safeguards | AI labs test models for dangerous capabilities and block harmful requests | Voluntary company frameworks, such as those published by OpenAI and Anthropic |
| DNA synthesis screening | Companies that make synthetic DNA check orders and customers before shipping | Industry protocol run by the International Gene Synthesis Consortium; a US mandate has been proposed |
| Government assessment | Intelligence and safety bodies judge which threats are real | MI5 centre launched in 2026; works with the UK AISI |
| Law and funding rules | Screening required for federally funded US research | Set by the 2024 US framework, which is still in force |
DNA screening is the option getting the most attention, because it acts at the point where digital ideas become physical material. In June 2026, the chiefs of OpenAI, Google DeepMind, Anthropic and Microsoft AI signed an open letter, organised by the Foundation for American Innovation and the Institute for Progress, asking Congress to make that screening mandatory.
A bill already exists. S.3741, the Biosecurity Modernization and Innovation Act of 2026, was introduced on 29 January 2026. It would have the US Commerce Secretary require providers to screen both orders and customers.
Where does US policy stand?
It’s patchy. A Congressional Research Service brief updated in August 2026 says no federal law gives anyone broad authority to regulate AI, including its use in the life sciences. A May 2025 executive order told the White House to revise the 2024 screening framework. As of mid-2026, that framework had not been replaced.
My view: the gap is understandable but hard to defend. Screening is cheap compared with the cost of a single failure, and the AI companies themselves are asking to be regulated here. When an industry asks for rules, lawmakers rarely get an easier opening.
What happens next
- The MI5 centre’s first outputs. Watch for whether its assessments reach AI labs directly, and whether any lab pauses a release because of one.
- US screening law. S.3741 and H.R. 3029 are both in the 119th Congress. If either passes, DNA screening would become a legal duty rather than a voluntary one.
- A revised US framework. The White House still owes the update the 2025 executive order asked for.
- Model testing. Expect more pre-release biology evaluations from AISI and its US counterpart as new models arrive.
Governments are also tightening their grip on AI tools for security reasons in other ways. Our piece on why DHS is pulling DeepSeek from US agencies is one example. Research on how “forgotten” AI knowledge can come back shows why removing dangerous knowledge from a model is harder than it sounds.
FAQ
Has AI been used to make a bioweapon?
Not as far as anyone has said publicly. McCallum described AI lowering the barrier to bioweapons as the misuse he weighs most heavily, but one that has not yet happened.
What is DNA synthesis screening?
It’s a check by companies that manufacture custom DNA. They review each order and customer before shipping, so dangerous material doesn’t reach the wrong people. Today it is largely voluntary.
Who is Tom Inglesby?
He directs the Johns Hopkins Center for Health Security at the Bloomberg School of Public Health. He has testified to the US Senate on the risks where AI and biology meet.
Want the full conversation? Listen to Ian Sample’s interview with Inglesby on the Guardian’s Science Weekly, then follow S.3741 to see whether screening becomes law.
Leave a comment