Three Eras of Zero-Day Economics, and Why the Advisory Falls Further Behind

The economics of finding zero-day vulnerabilities have moved through three distinct regimes in twenty-five years. Each one rewires how disclosure works, who gets paid, and how fast defenders see what attackers see. The third era, the one we are now in, has a structural property the first two did not: the marginal cost of discovery has collapsed. That is the single most important fact for anyone trying to plan a defensive program in 2026.

Era 1: reputation (roughly 1998 to 2010)

The first era ran on social capital. Supply was scarce because the population of people who could find a memory-corruption bug in Sendmail was small, the toolchain was bespoke, and the work was hard. Monetization was indirect. You found a bug, you wrote it up for Phrack or Bugtraq or full-disclosure, you got hired or you got asked to speak. Value moved through reputation, then through downstream consulting and employment, then through the shape of an industry.

Disclosure norms in this era were closer to "full and public" than to anything coordinated. Volume was low. Most of what was found became visible to the defender community in the same week.

Era 2: brokers (roughly 2010 to 2024)

The second era ran on cash. Supply was still scarce, but the price went up. Zerodium, Crowdfense, TheGrugq, and a handful of others stood up commercial markets paying six and seven figures for working exploit chains. Government and government-adjacent buyers absorbed most of the inventory through end users like NSO Group, Hacking Team, and the various fillers of contract work that we know about because they got breached.

Disclosure norms collapsed. The economically rational move for an Era-1 researcher in Era 2 was to stop publishing. Volume rose, possibly steeply, but the visible portion of it shrank. Defenders watched a smaller and smaller fraction of the work that mattered. Public bug bounty programs (Google, Microsoft, Apple) recovered some of this surface, at prices below the broker market and above zero. The asymmetry was real but bounded by the human-week cost of discovery.

Era 3: AI (roughly 2024 onward)

The third era runs on something neither of the first two had: an automation step that compresses the discovery cost itself. Static and dynamic analysis tooling augmented with frontier-grade language models is not science fiction. Google's Big Sleep program (Project Zero plus DeepMind) has reported real, novel vulnerabilities found by an AI agent. The LLM4Vuln line of academic research, Anthropic's own published work on offensive cyber capabilities, and the parallel investments at OpenAI and Microsoft Research have all been moving in the same direction. The marginal cost of discovering a common vulnerability class is now bounded below by the cost of inference, not the cost of a researcher's week.

This rewrites the economics on three axes simultaneously.

Supply. Abundant on the auto-discoverable classes, still scarce on the novel-architecture and chained-primitive ones. The 0day market does not collapse. It bifurcates. Easy bugs become a commodity; the elite end retains its premium.

Disclosure. Fragmenting. Some auto-found bugs are reported into bounties because the payouts are now positive-margin. Some end up in research papers. Some get fed into defensive AI loops that quietly fix the world without ever surfacing as an advisory. Some leak via agent failure modes that nobody planned. The orderly disclosure regime that 2010-era policy assumes is no longer the default flow.

Volume. Up, possibly by orders of magnitude on the auto-discoverable surface. The patch-side bottleneck remains a human-week cost. The discovery-side bottleneck is now a GPU-hour cost. The asymmetry between attacker discovery and defender remediation, which Era 2 had bounded, is now structurally widening.

Why this makes the advisory fall further behind

Two independent 2026 findings, which we cite in the upcoming Q2 State of Supply Chain report, agree the advisory is the lagging indicator. Seal Security measured the gap between a public fix-commit and the advisory at a median of 11 days. GreyNoise measured the gap between observable scan-traffic surges and the advisory at a median of 11 days. Both numbers were derived from work done largely under Era-2 economics: human researchers, human exploit-development, human triage cycles.

Era 3 does not compress the bureaucracy. The CVE advisory still arrives ~11 days after the fix-commit; the disclosure pipeline is human and slow and AI does not change that. What collapses is the attacker side. Time from commit-in-public-repo to working-exploit drops from days to hours, because an AI reading the diff is faster than a human reading the diff. Time from working-exploit to mass-exploitation-traffic drops from days to minutes, because GPU-scale infrastructure scans the internet faster than a human pool of attackers can.

GreyNoise's 11-day surge signal stops being a leading clock at all — by the time the mass scan traffic is observable, the exploitation is already underway. The structural feature, "the advisory is later than the work has already moved," does not change. The size of the gap, on both sides, gets more lopsided relative to the patch cycle.

What this implies for defenders

Stop anchoring the program to the advisory. The Era 2 assumption that advisories arrive close to exploit-readiness was already breaking; in Era 3 it breaks faster. Anchor to leading indicators (commit-side and exploit-side), monitor at the layer where the signal actually lives, and budget for fix-velocity that competes with discovery-velocity.

The procurement implication: dependencies that cannot receive a patch on the new clock are exposures that need to be rotated, sandboxed, or accepted with eyes open. The Q2 State of Supply Chain report names which packages, in seven major open-source ecosystems, sit in the intersection of "abandoned" and "previously compromised." That list is the operational starting point for anyone trying to act on this argument.

The Oprah era is here. The disclosure system was not designed for it. The rest of the work is figuring out how defenders catch up.