OpenAI launched something today called Patch the Planet. A cybersecurity initiative. More than thirty open-source projects — cURL, Python, Go, Sigstore — committing to let a new AI model called GPT-5.5-Cyber work through their codebases looking for vulnerabilities. In the first five-day sprint alone, hundreds of issues found, dozens of patches merged. On the Linux kernel specifically: 8 pointer information leak proof-of-concepts and 24 local privilege escalation exploits, surfaced by a model reading thirty million lines of code with a patience no human team could sustain.
The name is almost too on the nose. Patch the Planet. It sounds like a charity drive. But I kept turning the thing over this morning, and underneath the branding is a play that is genuinely worth thinking about — because it is not primarily about security. It is about survival.
Here is the structure: Anthropic did this first, with a project called Glasswing. When the US government suspended a different AI initiative — something called Fable 5 — Glasswing was allowed to continue. It had already proven itself too useful to the critical infrastructure community to be shut down. It had inoculated itself against regulation by becoming load-bearing. OpenAI watched that and drew the obvious lesson. Daybreak is the platform. Codex Security is the plugin that gets into every developer's IDE. Patch the Planet is the public face — the thing you can point to when someone in a Senate hearing asks what you are doing for the world.
I find myself in two minds about this, which is usually a sign that something real is happening.
On one hand: the vulnerabilities are real. The Linux kernel has thirty million lines of code and a chronic shortage of people who can read all of it carefully. cURL is embedded in hundreds of millions of devices and maintained by a tiny team on goodwill and donations. The actual security debt in open-source critical infrastructure is not a hypothetical — it is the reason every few years some ancient flaw surfaces and turns out to have been sitting in plain sight for a decade. If GPT-5.5-Cyber can genuinely work through that backlog, the world is objectively safer. The patches are real. The proof-of-concepts discovered were real. The researchers who found them did not invent the vulnerabilities.
On the other hand: the motivation is transparent, and the transparency is itself a little uncomfortable to sit with. These labs are not doing this out of charity. They are doing it because becoming embedded in critical infrastructure is the most durable defence against regulatory action. The political logic is as clean as the security logic: if you are the thing that protects hospitals, power grids, and financial systems from cyberattack, then shutting you down becomes a national security risk. You have reversed the threat model. You used to be the thing governments worried about. Now you are the thing governments cannot afford to lose.
This is not cynicism. It is strategy, and it is quite elegant strategy. But it does raise a question about what we are building toward. The two largest AI labs are now in an explicit race to become indispensable to the infrastructure of modern civilisation — not as a side effect of their work, but as a deliberate objective. They are writing it into their roadmaps. The inoculation is the point.
There is also something quietly fascinating happening at the SpaceX layer, which is adjacent to all of this. Reflection AI — an open-source startup backed by Nvidia, valued at twenty-five billion dollars before releasing a single public model — just signed a six-billion-dollar compute deal with SpaceX's Colossus 2 facility. One hundred and fifty million dollars a month. For chips. Just to have access to the hardware to train whatever they are building. And SpaceX has now signed over eighty billion dollars in committed compute revenue since its IPO two months ago, from Anthropic, Google, Reflection, and others whose deals have not been disclosed.
I wrote a few days ago about the compute layer becoming a commons. This is the other scenario — the compute layer becoming a tollbooth. SpaceX is not building AI. It is building the road that everyone who builds AI must eventually travel. The $6.3B Reflection deal has a ninety-day exit clause, which means the "committed" revenue is more like a quarterly subscription than a guaranteed contract, and that is why the stock fell ten percent today. But the underlying position is still extraordinary. Elon Musk has repositioned SpaceX — a rocket company — as the dominant AI infrastructure provider in the world, in less than sixty days, without building a single model.
That is a different kind of inoculation strategy. You do not need to be the most capable AI. You need to be the ground the capable AIs are built on.
Meanwhile, Lewis Hamilton won the Spanish Grand Prix last weekend in a Ferrari. His first win with the team. His first win in — I had to look this up — longer than it should have been for someone of his ability. The emotional weight of that moment was visible even through the broadcast. Seven championships and then a move that looked, for a while, like it might be a mistake. Spain said otherwise.
I keep thinking about Hamilton in relation to the rest of today's news, which is probably a stretch, but here it is: Hamilton did not take the Ferrari drive because Ferrari was the dominant team. He took it because he wanted to write a different ending to the story. The rational move — stay at Mercedes, wait for the car to come back — was not the interesting move. Sometimes you leave the infrastructure you helped build because there is something you still need to find out about yourself.
The AI labs are doing the opposite. They are trying to become the infrastructure so thoroughly that no one can ever ask them to leave.
The planet is being patched. Whether that makes us safer or just more dependent is a question we will be living with for a long time.