Source: Zero Day Clock: https://zerodayclock.com/
The clock tracks Time-to-Exploit (TTE) the gap between a CVE's public disclosure and its first confirmed in-the-wild exploitation. The trend is unambiguous:
- Mean time-to-exploit has collapsed from roughly 2.3 years in the earliest cohorts to about three weeks by 2025. In 2026 it reads in hours and keeps crossing zero, so it has stopped being a countdown at all. (Figure moves daily)
- True zero-days, CVEs exploited on or before disclosure day, have climbed from under 20% toward ~80% in the most recent cohort.
- The one-year window disappeared around 2021 and the one-month window around 2025; the one-week / one-day thresholds fall in 2026, with a one-hour window projected for 2027.
(The pre-2026 figures are observed and settled; the current-year readings will keep moving as the data matures, but the direction is unmistakable: disclosure and exploitation now happen almost together.
When exploitation coincides with disclosure, "read the advisory, then react" stops working. You can't out-hire or out-hour a speed problem. Defense has to move onto a different footing: continuously perceive risk across the whole environment, reason about what actually matters, and act at machine speed, with humans setting direction and keeping control of the consequential calls.
That splits the problem in two. Upstream, you find and fix the vulnerabilities in code you own before they're ever published, so the bug never becomes a public CVE, and there's no disclosure-to-exploitation race to lose. Downstream, for everything you don't control, the third-party CVEs the clock is really measuring, you can't out-patch the window any more, so you have to go looking. That means thinking like the attacker continuously; probing your own estate for the paths that actually lead somewhere, deciding which of them genuinely matter, and closing them at the same tempo attackers now operate. Neither is something humans can sustain at the required speed alone; both are things a coordinated system of agents can.
This isn't hypothetical. Microsoft's "codename MDASH" (Defense at AI speed: Microsoft’s new multi-model agentic security system tops leading industry benchmark | Microsoft Security Blog) is one public example of the upstream half, over a hundred specialized agents that discover, debate and prove exploitable bugs end-to-end, before disclosure.
Project Perception (Rethinking security for the age of AI - The Official Microsoft Blog) is an example of the downstream half. It's built as three teams of agents rather than one: red team agents that identify paths to compromise before an attacker can exploit them, blue team agents that investigate and reason over context to work out what represents meaningful risk, and green team agents that take corrective action and strengthen defenses across the environment. A closed loop of discover, evaluate, improve; running continuously, with humans in control of every consequential action.
The lesson underneath is worth holding onto, the advantage doesn't come from any single model or tool, but from the system and the discipline built around it. Chasing the newest model is the wrong reflex; building the loop that can perceive, reason, and act is the right one.
Put it together and the picture is clear. The Zero Day Clock quantifies the threat; agentic, machine-speed defense is the shape of the response. Human-speed discovery and triage are now structurally outpaced, which doesn't make security teams obsolete, it changes what we're for: setting the strategy and the guardrails, and keeping humans in control while the machines carry the load.
Whether organisations respond through agentic systems, automation, or other approaches, the underlying challenge is the same: security operations must increasingly operate at machine speed, while humans remain responsible for strategy, governance, and consequential decisions.
Disclosure: I work at Microsoft as a Cloud Solution Architect. The views expressed here are my own and all referenced sources are publicly available.
]]>