Assume everything is exploitable: vulnerability management after AI

The frame

For twenty years, vulnerability management operated on a reasonable premise: there are more vulnerabilities than you can patch, so you prioritize. You look at CVSS. You look at whether exploit code exists in the wild. You look at whether the asset is internet-facing. You make a call.

What changed is not the volume of vulnerabilities. It is the cost to exploit them.

Claude Mythos, in its most recent iteration, can autonomously discover zero-day vulnerabilities and build working exploits for well over half of what it finds. It chains vulnerabilities together. It builds multi-step attacks. It does this across every major operating system and every major browser, including obscure bugs that have sat dormant for twenty years.

This is not a projection. Anthropic runs Project Glasswing, giving forty-plus companies early access to the vulnerabilities Mythos discovers, specifically so they can patch before adversaries get the same capability. OpenAI runs a parallel program.

The capability exists. The cost curve for exploit development has collapsed. The assumption that a vulnerability without public exploit code is a lower priority is no longer defensible.

The floor rose for everyone

AI did not just accelerate the most sophisticated adversaries. It democratized attack capability across every tier.

Hacktivists, historically the least capable actors, now operate at a sophistication level that belonged to nation-states three years ago. Cybercrime syndicates gained capabilities that previously required elite teams. Insider threats, someone with an axe to grind and legitimate access, can now augment their attack with AI-generated exploits they could not have built themselves.

The nation-state tier, already dangerous, is now operating at a level that has no precedent. They are fully supercharged.

The implication for vulnerability management is straightforward. You are not defending against the adversary you planned for. You are defending against every adversary operating one or two tiers above where they were when your program was designed. The program has to close the gap.

The assumption that has to change

The old prioritization question was: is this vulnerability being actively exploited?

The new question has to be: what happens if it is?

When exploit development is scarce, the Known Exploited Vulnerabilities list is a meaningful filter. When AI can generate exploits autonomously, it becomes a lagging indicator. The vulnerability that shows up on KEV tomorrow was exploitable last week. The vulnerability that never shows up may have been exploited silently.

This does not mean you stop using KEV or CVSS. It means you stop using them as the primary filter. They become one input among several, and the ordering question changes from "which vulnerabilities are most dangerous in the abstract" to "which vulnerabilities sit on attack paths to assets that would be existential if compromised."

That is a different question. It requires knowing what your crown jewels are, not just what your scanner found.

What the proactive SOC actually means

Most security organizations spend eighty percent or more of their resources on the right side of the NIST framework: detect, respond, recover. The work of monitoring, triaging alerts, and running incidents to ground.

That ratio has to invert. The proactive side, identify & protect, has to become the primary investment.

The workflow is not complicated, but it is sequenced: discover what you have (assets, vulnerabilities, identities, human and machine), synthesize the data into a single view (deduplicate across tools, identify coverage gaps where assets lack scanners or EDR), assess what is critical and what mitigating controls exist, then secure through patching and control remediation.

The step most organizations skip is the synthesis layer. They have asset data in one tool, vulnerability data in another, identity data in a third, and no unified view of which critical assets have which vulnerabilities behind which controls. That gap is where AI-powered attacks will enter.

What this is not

This is not a tool recommendation. The tools are downstream. This is not a model comparison. The models will change, and the gap between Mythos today and GPT tomorrow will close faster than most planning cycles can respond. This is a briefing on the assumption that has to change before any tool or model decision can be made intelligently. Change the assumption, and the rest of the program can be re-sequenced around it. Keep the old assumption, and every investment downstream is calibrated to a threat model that no longer applies.