The patch-and-pray era is ending.
October 2026 has delivered an extraordinary cluster of critical security disclosures, including flaws in foundational internet infrastructure. Some are calling it “Black October.” Others call it the start of the “Vulnpocalypse.” Regardless of the label, AI-accelerated discovery is surfacing bugs faster than anyone can fix them, and the exploitation windows are shrinking from weeks to hours.
Sam Altman says he no longer thinks we can simply patch all the bugs, and he argues leaders haven’t grasped how exponential this is. Anthropic’s Deputy CISO Jason Clinton warns of a “vulnerability tsunami” and says things will get much worse before they improve. Former NSA cybersecurity director Rob Joyce describes a shift from vulnpocalypse (mass discovery) to breachpocalypse, where AI agents can jiggle every doorknob, continuously, while humans cannot keep up, especially on legacy and unsupported systems.
As for how developers are asked to deal with it, the recommendations do rhyme. Prioritize patching internet-facing and critical systems. Be ready to take them offline if you can’t patch them quickly. Isolate what you can’t patch. Audit continuously with AI and disclose responsibly.
All of this is reasonable. They are also, I think, insufficient.
The optimism problem: “vulnzero”
Clinton’s long-term answer is what he calls vulnzero. His argument, posted on X around October 22, is that scaling principles will push mythos-class cyber capability into small, cheap models that developers can run continuously.
That is a genuine possibility, and I hope he’s right. But it rests on an assumption that cheap, ubiquitous AI helps defenders more than attackers. I doubt it.
The oldest asymmetry in security is that it is easier to create and find bugs than to catch them. An attacker needs one exploitable weakness. A defender must maintain security across every dependency, configuration, and piece of operational infrastructure. With new tech, new attack surfaces will always emerge, so faster, AI-powered patching will not completely fix it.
A different premise
I no longer believe in the open web as we’ve known it.
The default assumption behind the internet, that publicly exposed systems accept requests from unknown actors and then decide how to respond, only works for an era when probing was expensive and attackers were scarce. An autonomous adversary can discover services, fingerprint technologies, read public code, test interfaces, and adapt, all at scale and nearly for free.
So my argument is that two fundamental assumptions deserve to be retired:
Access by default. An unknown actor should not be able to interact with protected functionality without first establishing some form of trust.
Identity as an account. An application should not issue digital identity to you. You should control it and use it to establish your own authority.
What I think should remain is user autonomy, and the accountability that comes with it. Autonomy means people control and own their identities, credentials, data, software, and the agents they delegate to. Accountability means those people, or the systems they authorize, face meaningful constraints and consequences for what they do.
The industry’s current answer to the AI-driven vulnerability surge is to triage harder, isolate more, and wait for cheap defensive models to catch up. And while those are necessary, they don’t change the fact that anyone, or anything, can walk up to a system and start probing. And that once inside, there’s little to no mechanism to implement accountability.
I’m not saying let’s go for a surveillance machine. There are far better ways to do it without logging everyone’s online identities and actions.
The permissionless Internet
The internet should remain permissionless to participate in, but authority to act upon other people’s systems and resources must become explicit, verifiable, and accountable.
We will not out-patch an attacker that never sleeps. This is a domain for defensive technologies and not for another shiny AI model to fight asymmetric battles. When something does go wrong, how much damage can it actually do?
Clinton may be right that in 18 months or so the flood of shipped bugs recedes. But the transition suggests that we are forced to rebuild anyway. We should build internet systems where users retain control over their identities and authority, where we don’t need to evade user privacy to establish trust, and where compromising one component does not entail compromising everything. We are having an increasingly adversarial digital environment, and the challenge is how to make every consequence manageable and resilient in case something breaks.