24 October 2026
ITV News's 'UK to Set Up AI Biosecurity Lab' Announces a First Without Answering the First Question
Nobody has written down what OpenAI can do with a pathogen signal it just detected for the British state.
By end of 2027, the AIBL will publish a rule naming what its model providers can do with biological threat data they process. If there's no rule by then, the arrangement runs on trust. That matters, because trust already failed this month.
Chancellor John Healey announced the AI Biosecurity Lab on 24 October1, backed by £100m in public money and up to £200m in credits from OpenAI and Anthropic. The budget comes next week. Good. Biological threats are real and AI detection could help. The lab should exist.
Here is what the announcement doesn't answer.
1. The £200m is rent. Credits mean the state pays to run queries on servers it doesn't own. Two-thirds of the lab's announced money comes from the two companies whose models will read its data.
2. The data is the sensitive thing. To spot a biological threat with AI, you describe the threat to the model. Genomic sequences. An outbreak in a city before it's public. A pathogen nobody has named yet. That's not a web search. It's the exact intelligence an enemy state would want, sent to servers the British state doesn't control.
3. Nobody has published what happens next. Not what the model providers can't train on. Not what they must delete. Not who checks. The AIBL announcement says nothing about it. The Biological Security Strategy2 was written before the question existed. The AI Security Institute's rules cover the state looking at a model, not a model looking at the state's data.
4. First means no one else has done it. There's no comparable deal to copy, no agreed norm for how state biosurveillance data flows through a commercial model. Britain is making the rules up as it goes, without writing them down.
5. This is last month's lesson, run backwards. Earlier this month3 the AI Security Institute lost access to frontier models for two and a half weeks when an arrangement froze. That showed the whole evaluation programme sat on consent, not law. The AIBL starts the same way. But the data runs the other way. Not the state looking at a model. A model looking at the state's most sensitive files.
Yes, the budget is next week. The rule may still come. But the government named OpenAI and Anthropic's credits as the financial bedrock and called this a world first three times in a day. You don't announce what you're building and then decide who owns what it sees.
The state kept the power to quarantine a pathogen. It gave a commercial model the first look at it.