AI Policy

More Than 15 Partners, and No Published Rules

Google DeepMind says it advanced more than 15 biosecurity partnerships in 12 months and named none of them. There are no published criteria, no application process and no appeal path. That is a procurement fact, not an ethics complaint.

AI PolicyBy Servola Tech Desk2026-07-176 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

More Than 15 Partners, and No Published Rules

Key takeaways

Google DeepMind says it has advanced more than 15 biosecurity partnerships in the past 12 months and it names none of them. Access to its most powerful biology models is granted at the company's discretion, and no eligibility criteria, no application process and no appeal path has been published.
The safety case for gating these models is strong and we are not disputing it. Our point is narrower: unpublished rules are unplannable rules, and that makes this a procurement question.
The wider programme does name collaborators - Lawrence Livermore National Laboratory, the UK AI Security Institute, CEPI and the Francis Crick Institute. Two of those are British institutions, so the gate is already operating inside UK research.
Demis Hassabis is publicly pushing for a government frontier-AI standards body while his own company runs a private one. Read that as the most honest thing in the announcement: today's arrangement is a placeholder nobody is comfortable with.
For every AI capability in your stack, write down who can withdraw it, on what stated criteria, and what your appeal is. Where any of those answers is "not published", you have found a single point of failure you have not priced.

Fifteen partnerships, and a list that is not there

On 16 July 2026, Google DeepMind published a post called "Our approach to bioresilience", written with Isomorphic Labs. Most of it reads like science. The sentence that matters is administrative. "We are making our AI models and agents available to trusted partners," it says, across three pillars: prevention, detection and response. The technologies named around that sentence are the company's best work: AlphaFold, AlphaGenome, Gemini, AlphaEvolve, SynthID, and Isomorphic Labs' Drug Design Engine.

Then comes the number. "Over the past 12 months, we have advanced more than 15 partnerships" with government bodies, biosecurity organisations and research groups. Fifteen is a real figure, volunteered by the company, describing a year of work that has already happened. The post names none of those fifteen.

The silence is the story. There are sound reasons a biosecurity partner might prefer not to appear on a list, and we are not suggesting otherwise. The difficulty is that the number arrives with nothing a reader could use to work out who sits inside it. Fifteen partnerships have been advanced. Trusted partners have access. The post never defines trusted.

What we went looking for, and did not find

The honest finding here is a negative one. We read the post for the mechanics of access: who qualifies, how an organisation applies, who decides, and what happens when the answer is no. None of it is there. No published eligibility criteria. No application process. No stated appeal path. That is an absence we checked for and confirmed. DeepMind has not said anything objectionable on the subject. It has not said anything at all.

It is worth keeping whose words are whose straight. "Capability-gated access" is our phrase. It appears nowhere in the post, and DeepMind should not be quoted as having used it. Our read: a private company has built a gate in front of its most powerful biology models and has advanced more than fifteen relationships through it without publishing the rules the gate runs on. That characterisation is ours and it should be weighed as ours.

The safety rationale deserves saying plainly, because it is strong. Models that design proteins are dual-use in a way ordinary business software never was, and a company that handed them to whoever asked would be behaving badly. Helen King, DeepMind's vice president of responsibility, put the position to Axios: "If we were to find that we were reaching a critical capability level and we didn't have the appropriate mitigations, then we would not be launching." That is the right instinct, and it does not settle the point. Our objection is not to the gate. It is to a gate with no published rules.

Two of the names that are in the post are British

The post is not anonymous everywhere. Alongside the unnamed fifteen, DeepMind names collaborators in the wider bioresilience programme: Lawrence Livermore National Laboratory, the UK AI Security Institute, CEPI and the Francis Crick Institute. Two of those are British institutions. For a UK or European reader this is not a distant American story about American laboratories.

The work described around those names is concrete. AlphaEvolve is being applied to optimise metagenomic sequencing algorithms, with the aim of detecting an outbreak sooner. SynthID, the watermarking system most people met as a way of labelling AI-generated images, is being adapted to biology for DNA synthesis screening. Isomorphic Labs has stood up a dedicated unit for the rapid deployment of medical countermeasures. This is real infrastructure, and part of it is being built inside British institutions.

That sharpens the question rather than settling it. If the UK AI Security Institute and the Francis Crick Institute can be named, then DeepMind is plainly able to name a partner when it chooses to. The choice is being exercised more than fifteen times over, and nothing published tells a reader on what basis.

Hassabis is asking for the gate he already operates

The most interesting thing in the announcement is a contradiction, and it is not hypocrisy. Axios reported that Demis Hassabis is pushing for a government body to set frontier AI standards, and that DeepMind's Owen Larter pointed to cross-lab agreement on pre-release testing. So the company running a private, unpublished access gate is publicly asking someone else to build a public one with rules.

Our read: that is the most honest thing in the whole announcement. A company convinced its own discretionary gate was the right permanent arrangement would not lobby for a statutory alternative. Asking for a standards body is an admission that today's setup is a placeholder, held in position by the judgement of the people who happen to hold the posts.

Placeholders are useful. Depending on one without noticing that is what it is, is the problem. Helen King's line about not launching without appropriate mitigations is a statement of intent from a named executive at a named company on a given day in July. It binds nobody once she leaves the role, and no owner can hold it up in a dispute.

A permission you have booked as a capability

Here is the part that leaves biology behind. A capability your supplier grants at its own discretion is a permission. The two look identical on a system diagram and behave differently the moment something goes wrong, because a capability has terms, a price, a notice period and a remedy, and a permission has a person at the other end who is entitled to change their mind.

So the instruction is paperwork, and it will take an afternoon. For each AI capability in your stack, write down three things: who can withdraw it, on what stated criteria, and what your appeal is. Published, not assumed, and not what an account manager said on a call. Where any of the three answers is "not published", you have located a single point of failure you have never priced, because nothing on your invoice reflects the chance that it goes away.

This is not a biology problem. The same gate shape is arriving across frontier model access generally, wherever the safety case and the commercial lever turn out to be the same lever in the same hand. DeepMind's bioresilience post is the clearest published example so far, and it is clear only because the company was candid enough to put the number in writing.

Frequently asked questions

Is Google DeepMind doing something wrong here?

No, and we are not arguing that it is. Gating powerful biology models is a defensible safety decision and the rationale for it is strong. Our point is narrower. The rules of the gate are unpublished, and rules you cannot read are rules you cannot plan around.

Which fifteen partners are they?

The post does not say. It states that more than 15 partnerships have been advanced over the past 12 months, with government bodies, biosecurity organisations and research groups, and it names none of them. The wider programme does name Lawrence Livermore National Laboratory, the UK AI Security Institute, CEPI and the Francis Crick Institute, but those are not presented as the fifteen.

Did DeepMind call this capability-gated access?

No. That phrase is our characterisation and it appears nowhere in the post. What the post says is that the company is making its AI models and agents available to trusted partners. It does not define trusted.

My business has nothing to do with biology. Why does this matter?

Because the mechanism is not biological. A supplier deciding at its own discretion who may use a capability, on criteria it has not published, is a pattern arriving across frontier AI access generally. The exercise of writing down who can withdraw each capability in your stack applies whatever that capability does.

What should I actually do this week?

List every AI capability your operation depends on. Against each one write who can withdraw it, the published criteria they would use, and your appeal route. Every "not published" is an unpriced single point of failure, and the finished list is usually shorter and more uncomfortable than owners expect.

Fifteen partnerships is a disclosure. The missing list is a governance fact, and it is the shape of what is arriving. The companies building frontier capability are also, for now, the ones deciding who may hold it, and they are doing it without publishing the rules. An owner does not need an opinion on whether that is right. An owner needs to know which of their own dependencies sit on the far side of a door somebody else controls.

Google DeepMindIsomorphic LabsBiosecurityAI governanceModel accessUK AI Security Institute

The Prompt Running Your Business Has No Owner

On 9 July 2026 Mistral put prompts and skills under version control, ownership and audit logs. The real story: the instruction has become your software, and most European firms run it ungoverned.

4 min read

Artificial Intelligence

Your AI Bill No Longer Has a Ceiling

AI tools are moving from fixed per-seat pricing to metered token billing with no natural ceiling. Here is how owners keep AI spend under control.

3 min read

Digital Sovereignty

The EU Council Revived the Rule That Scans Private Messages

The EU Council revived the lapsed rule that lets providers scan private messages with AI. For owners, confidential communication is now a governance choice.

2 min read

← All articles