Discovery Is Now a Commodity. Closure Is Not.

What happened in that week

Google launched Fairwind and gated Gemini 3.8 Flash Cyber to vetted governments, critical infrastructure operators and core technology platforms, scoping the model to vulnerability detection and automated patching. Anthropic shipped Fable 5.1 and Mythos 5.1 with tiered safeguards, on top of the Glasswing coalition it had already assembled. OpenAI published its own framing on putting frontier cyber models in more trusted hands.

Three labs, three access-control regimes, one week.

The access controls are the headline for most readers and they are not the interesting part. Gated access is a decision about timing, and timing decisions have a short half-life. Capability diffuses; it always has. The question worth asking is not who gets the model in September but what the world looks like once everyone has something like it — because that is the world we are building for.

The interesting part is what Google DeepMind said about why it built what it built: it prioritised fixing over exploitation, deliberately. That is a remarkable thing for the supply side of a market to say about its own product.

The paragraph I had to delete

My original draft went from there to a victory lap. The organisations building the discovery capability — the ones with every commercial incentive to tell you discovery is the valuable half — have independently concluded that the fix side is where the scarce value sits. Therefore the debate about where the constraint lives is over. Therefore we were early.

That paragraph did not survive contact with the market.

In March, Surf AI launched out of stealth with $57 million, led by Accel, on the explicit premise that the industry has acknowledged but not solved the inability to reliably execute remediation at scale. Gomboc has raised on the argument that defensible remediation is now a bigger problem than detection, and has been saying so to regulated buyers for a while. Neither of those is a small company with a clever tagline. Both are serious teams with serious backing making a version of the argument I thought I was making against the grain.

So the honest position is not that the labs settled the argument. It is that the argument was settled before the labs joined it, and I should say so rather than claim a consensus as a prediction.

I think that matters for a reason beyond intellectual hygiene. If a thesis is only interesting while it is contrarian, it was never a thesis. It was a marketing position. The useful question is what remains true once everyone agrees with you.

Agreeing on the diagnosis is not agreeing on the architecture

Here is what remains true.

“Execution is the constraint” is now a crowded sentence. “Here is what execution safely requires” is still an empty room.

Almost every version of the closure argument currently in market stops at orchestration: connect the systems, route the work, let agents act on policy, close more things faster. That is a real capability and it produces real results. It is also the easier half, and it quietly assumes the finding was worth acting on in the first place.

That assumption is where the risk lives. A closure engine is a multiplier, and a multiplier applied to an unvalidated input stream multiplies the error rate along with the throughput. Automate on top of a discovery layer that over-rates severity and you do not get more security — you get more change requests, more regression surface, more emergency windows, and you spend the organisation’s willingness to let a machine touch production on the first bad push.

Speed is not the scarce resource. Warranted speed is.

Not all closure is equally hard, and it is worth being precise about why

I want to be fair to the other companies working on this, because the distinction between what they have solved and what remains is technical rather than promotional.

Surf’s domain is identity and operational hygiene — dormant and orphaned accounts, access governance, unused licences, configuration risk. Gomboc’s is infrastructure-as-code misconfiguration, Terraform and CloudFormation, policy-driven enforcement. Both have picked domains where automated closure is genuinely tractable, and the reason is the same in both cases.

In those domains, correct is declarable in advance. A dormant account should be disabled. A security group should not be open to the world. A bucket should not be public. The desired state can be asserted as policy, which means the fix can be verified against it by construction. You do not need to establish whether the fix was right; you need to confirm the resource now matches the policy. That is a checkable property, and building a deterministic engine against a checkable property is good engineering.

A vulnerability in application code or a transitive dependency does not work that way. There is no policy statement that tells you whether the vulnerable path is reachable in your deployed configuration, what privilege the affected process actually runs with, whether the available fix crosses a breaking API change, or whether the regression surface is acceptable for that service this week. Correctness is not declarable ahead of time. It has to be established — per finding, per environment, every time.

That is why these domains commoditised in the order they did, and it is why the hardest part of the closure problem is still in front of the industry rather than behind it.

What deterministic has to mean

The word is getting loose, and I would rather be precise than territorial about it.

Gomboc uses deterministic to describe fix generation: the same input produces the same Infrastructure-as-Code output, rather than whatever a generative model feels like emitting today. That is a correct and valuable use of the term for that problem, and the contrast they draw with generative unpredictability is a fair one.

Our use of it sits one step upstream, and it is a different step. Deterministic validation — establishing that a finding is real, that the vulnerable path is reachable, that severity holds in this specific system, and doing it the same way every time — before anything downstream is authorised to act. Same input, same verdict, always, with evidence a reviewer or an auditor can check without taking the reasoning layer’s word for it.

Both steps are necessary. Only one of them is widely built. A deterministic fix generator sitting behind a probabilistic severity judgment is a very reliable way to ship the wrong change quickly.

You cannot win a machine-speed war with human-speed defences. You also cannot win it by removing the human and keeping the guesswork.

Throughput is the product

If discovery is an input you rent, the thing worth building is the machine that turns findings into closed risk without a human in every loop — and that machine is mostly not a model.

It is a validation layer that establishes what is real and what matters, deterministically, so everything downstream has something trustworthy to act on. It is a remediation path that can generate, test and stage a fix. It is an integration surface into the systems where change actually happens. It is compensating control that holds the exposure while a fix moves through a change window it cannot skip. And it is an evidence trail that survives an auditor who does not care how impressive the model was.

None of that commoditises, because it is entangled with the specific mess of a specific enterprise. That is exactly why it is hard, and exactly why it is worth building rather than bolting onto the same fragmented stack.

Our own research says 65% of organisations still run vulnerability management at least half manually, and 60% need more than a week to remediate a critical. Those numbers do not move because a better model became available. They move when the pipeline changes shape.

What I would ask now

Not “which model do you use.” That answer will be obsolete in two quarters and converged across the market in four.

Not “do you believe closure is the constraint,” because everyone now says yes and the answer costs nothing.

Ask what sits between the finding and the fix. Ask what the system does when the finding is wrong — not whether that happens, but what happens when it does. Ask what fraction closes without a human, and what holds the exposure for the fraction that cannot. Ask what evidence comes out the other end, and whether you could hand it to a regulator without a translator.

The labs have told you where the scarce resource is. So has the market. Discovery is becoming something you rent, and closure is something you have to build — but the part nobody is building is the layer that decides whether closing was the right thing to do at all.

‍