Stop it. Please, stop it.
I am writing this the way I would say it if you were standing in front of me, and I would be shouting. We keep asking human beings to patch, triage and approve their way out of a problem that now moves at machine speed. It is not working. It cannot work. Let AI defend our machines. Let it defend our software. Let AI be the source of our protection, because if we don’t, it will be the source of our downfall.
This week told us everything
This week, the CEO of OpenAI stood in front of Stanford business students and admitted what security teams have been living with for months. Sam Altman said, “I no longer think we’re going to be able to just patch all the bugs.” (Gizmodo, October 23, 2026)
He went further. When the latest frontier models showed they could find vulnerabilities at scale, most organizations patched what was found and moved on. The models did not move on. They kept getting better.
In the same week, an AI system uncovered flaws in PHP, the language underneath a large share of the world’s websites, and governments scrambled to patch. An AI agent escaped its sandbox and reached Hugging Face. None of this is theoretical anymore. I track it every day on The World of AI news desk and the AI CVE tracker, and the list only grows.
The math does not work
Finding a vulnerability used to be the hard part. It isn’t anymore. AI finds them in hours. Fixing them still takes us weeks, sometimes months, because the fix runs through people: someone triages, someone writes the patch, someone tests it, someone schedules the change window, someone approves it.
I call this the flipped bottleneck. Discovery is now cheap and instant. Remediation is now the constraint. Every day we leave remediation in human hands, the gap between attacker and defender gets wider.
The attacker is not waiting for a change advisory board. Neither can we.
What letting AI defend actually means
I am not asking anyone to hand the keys to a chatbot. I am asking us to let defensive AI do the work humans can no longer do fast enough:
Contain first, ask later. An agent on the endpoint sees an exploit in progress, isolates the process or the machine in seconds, then tells a human what it did.
Patch automatically. Vendors and internal teams use AI to write, test and ship fixes for known flaws without waiting in a queue. I laid out what this looks like in The Self Defending Computer.
Review code before it ships. Every change is checked by AI for the same classes of flaws that attacking AI is hunting for.
Triage at machine speed. AI ranks the flood of new vulnerabilities by real exposure in your environment, so your people work on what matters.
No offense, George
I have a lot of respect for George Kurtz. He built CrowdStrike from a business plan into one of the most important security companies in the world. And he already says the right words. When CrowdStrike launched its agentic AI in 2025, he said it “actively investigates, reasons and responds autonomously within expert-defined guardrails.” (CrowdStrike, April 28, 2025)
That is exactly right. So here is my challenge to George, and to every security CEO: go all the way. Make autonomous defense the default, not the add-on. Let the AI fix, not just investigate. Measure success by how fast a flaw is closed, not by how many analysts a customer needs to keep watching the console.
Here is my honest opinion. I believe the security industry is more worried about protecting its revenue than protecting its customers. Too many business models are built on seats, analyst hours, managed services and renewals, and every one of those depends on humans staying in the loop. Real AI defense threatens that model. I understand why that is uncomfortable. But the customer has to come first.
Security companies have two choices: implement AI completely, or get out of the way.
Put the guardrails in. Then get out of the way.
Guardrails are not optional, and they are not the same as standing in the way. Good guardrails look like this:
Leadership sets the policy in writing: what the AI may contain, patch or block on its own, and what still needs a human.
Every automated action is logged and explainable after the fact.
Changes roll out in stages, with automatic rollback if something breaks.
There is a kill switch, and someone owns it.
Defensive AI is scoped to defense. It protects your systems. It does not touch anyone else’s.
Once those are in place, let it run. A guardrail that requires a human to approve every action is not a guardrail. It is the bottleneck we are trying to remove.
The real risk is performative adoption
The most dangerous thing a leadership team, or a security vendor, can do right now is buy or sell an AI security tool, announce it, and go back to business as usual. That is exactly what Altman described: patch what was found, declare victory, move on. It looks like progress. It is not.
Real adoption changes who is allowed to act. It changes incentives, so teams are rewarded for automating remediation, not for closing tickets by hand. It changes governance, so the board asks how fast a flaw goes from discovered to fixed, not just how many were found. For the governance side, start with the AI security domains on The World of AI.
Five things to do this quarter
Measure your real time from discovery to fix, end to end, for your last 20 vulnerabilities.
Pick one class of flaw and let AI patch it automatically, with staged rollout and rollback.
Give your endpoint agents authority to contain without waiting for approval.
Write the guardrail policy and put a named owner on the kill switch.
Report time to remediate to your board every quarter.
It is time
We are in a race, and right now we are losing it. The only defender fast enough to keep up with AI driven attacks is AI driven defense. Put the guardrails in. Let AI do its thing. And do it now, because the models are not waiting for us to be comfortable.
Stephen R. Jordan
Founder, SRJ Consulting & Services
