Microsoft's own telemetry, not a vendor pitch deck, is the source for a specific and uncomfortable number: the median time from a vulnerability's public disclosure to its first real-world exploitation has fallen to well below 24 hours. The company's 2026 Digital Defense Report, covering July 2025 through June 2026, frames this as a direct consequence of attackers adopting AI faster than defenders have -- "AI is changing the physics of cybersecurity," in the report's own words.
The clearest evidence for that claim isn't abstract. In a controlled evaluation run separately from its telemetry collection, Microsoft had two frontier models -- Anthropic's Mythos Preview and OpenAI's GPT-5.5 -- chain 32 consecutive attack steps into a full domain compromise of an emulated enterprise network, with no human operator directing any individual step. Microsoft is careful to frame this as a controlled-environment result, not a report of an attack that has happened in the wild -- the gap between a lab demonstration and a real campaign is the whole reason the report calls this a warning rather than an incident.
The shift shows up in how attackers are actually getting in, not just in how fast. Phishing's share of initial access roughly tripled, from 7% to 23%, over the year Microsoft tracked, and exploitation of public-facing applications rose from 15% to 24% over the same period -- even as social engineering overall, a broader category, fell from 15% to 7%. Read together, that's a narrowing: attackers converging on a smaller number of AI-accelerated entry points rather than spreading effort across many tactics.
- Phishing as initial-access vector
- Public-facing application exploits
- Social engineering overall
The report's full-year vulnerability count adds scale to the speed figure above. Microsoft's own telemetry counted roughly 40,000 CVEs disclosed industry-wide in just the first half of 2026 -- on pace, the company says, to roughly double by year-end -- while Infosecurity Magazine's independent read of the same report put the full-year estimate at closer to 72,000, itself a record.
What each headline figure in Microsoft's report actually covers
- <24 hrs · median
- Vulnerability-to-weaponization time
Includes: The gap between a CVE's public disclosure and its first observed real-world exploitation, across Microsoft's full telemetry base
Excludes: Enterprise remediation time, which the report separately puts at 30-60 days for critical external vulnerabilities - ~72,000 · est. for 2026
- CVEs disclosed industry-wide
Includes: Publicly disclosed vulnerabilities tracked through the report's full-year sources
Excludes: Microsoft's own first-half count of roughly 40,000, which it separately projects would roughly double by year-end -- a close but not identical estimate - 32 · steps
- Longest chained autonomous attack Microsoft tested
Includes: One controlled-environment evaluation ending in full domain compromise
Excludes: Any confirmed real-world campaign reaching the same length -- Microsoft's wild-caught examples, like the JadePuffer campaign, are shorter and still partly human-directed
Microsoft's breakdown of who gets targeted adds a geography and sector dimension the top-line numbers don't: government agencies accounted for 27% of all attacks the company tracked, with IT at 17% and research and academic institutions at 14%. By country, the United States took the largest single share at 25.5%, followed by Israel at 7.6%, Ukraine at 4.8%, and Taiwan at 3.9% -- a list that tracks geopolitical flashpoints more closely than it tracks GDP.
- The single most-targeted sector, and the one where a successful breach carries the broadest downstream exposure.
- Together account for nearly a third of attacks tracked -- both sectors that sit upstream of many other organizations' own security.
- Microsoft's own comparison puts them roughly 29 to 59 days behind attackers using the same published vulnerability.
- The report doubles as a sales case for the AI-driven defensive tools it is simultaneously selling to the organizations it says are falling behind.
Not every tactic moved the same direction. Credential-based intrusions held steady at a high level -- 52.2% of valid-account intrusions involved additional credential theft, and 18.4% involved active password-spraying campaigns -- which Microsoft frames as evidence that old-fashioned credential abuse hasn't been displaced by AI-driven tactics so much as sped up alongside them. One incident the report cites by name, a malicious browser extension harvesting AI chat conversations, reached more than 600,000 installs and touched nearly 10,000 organizations before discovery.
Microsoft's own framing draws a careful line: none of this, the report says, represents a new attack category. The physics changed; the chemistry didn't is one way to read that distinction, and it matters for what the finding does and doesn't justify -- a faster version of a known problem argues for faster patching and detection, not for an entirely different defense model.
"While none of these represent new attack methods, the quantitative increase in the scale and speed of attacks creates an immediate problem for defenders."
- Microsoft's 2026 Digital Defense Report says vulnerabilities are weaponized in under 24 hours, far faster than patching.
- In a controlled test, two frontier models chained 32 attack steps into a full network compromise with no human direction.
- Phishing's share of intrusions roughly tripled to 23%; broader social engineering actually fell.
- Government agencies were the single most-targeted sector, at 27% of attacks Microsoft tracked.
- Caveat: the 32-step chain was a controlled lab test, not a confirmed real-world attack.