Attackers Now Move in a Day. Here Is an Update Routine That Keeps Up.
Microsoft and Google both say AI has shrunk the time between a flaw going public and an attack. You don't need a new product. You need a short list and a weekly habit.
By John Martines · October 9, 2026 · 6 min
Microsoft says the median time from a flaw being found to a working attack is now well under 24 hours, and Google says flaws found by AI are the dangerous kind. Here is the routine a small business can actually keep.
The firewall in the back closet has a small notice on its admin page: "Firmware update available." It has said that since August. Nobody minds, because the internet works and the update means a reboot during business hours.
That notice used to be a low-stakes chore. This month two of the biggest security teams in the world published numbers that change the math. When a serious flaw goes public now, attackers can be using it within a day.
A vulnerability is a flaw in software that lets someone do something they shouldn't, like get into a system without a password. When the maker finds out, it releases a fix (a patch or update) and publishes a notice. That notice is also a signal to attackers: here is a door, and here are the businesses that haven't locked it yet.
The gap between "the fix is out" and "you installed it" is your patch window. It is the time you are exposed.
Two reports this month say AI has squeezed that window from both ends.
Microsoft's 2026 Digital Defense Report, covering July 2025 to June 2026, says the median time from a flaw being found to a working attack is now well under 24 hours. Large companies, it says, typically take 30 to 60 days to fix critical flaws on systems that face the internet. It expects a record of about 72,000 publicly tracked flaws this year.
Google's Threat Intelligence Group found that monthly flaw disclosures roughly doubled between January and August 2026. Flaws that look to have been found by AI tools were more serious on average: half of them let an attacker run their own code on the target, compared with about a quarter of the rest.
How it got here:
- November 3, 2021: CISA, the US cybersecurity agency, started its Known Exploited Vulnerabilities catalog, a public list of flaws attackers are actually using, and urged businesses to watch it.
- February 6, 2026: BeyondTrust disclosed a critical flaw in its remote support software that an AI research agent had found. Google says one attack group was using it within four days, and five more within a week.
- September 30, 2026: Google published its findings on AI and vulnerability discovery, including more than 1,500 flaws disclosed this year in AI software itself.
- October 2, 2026: Apple said it will add tighter controls to Full Disk Access on the Mac, the setting that lets an app read everything on the computer, because the risk grows as AI agents become more capable. It did not give a date.
- October 2026: Microsoft's report landed with the under-a-day figure.
There is good news in the same reports. Google counted 141 flaws that were disclosed and then used by attackers in the first eight months of 2026. That is about 1 in every 431 disclosed flaws. You do not have to fix everything the same day. You have to fix the right few things fast.
The right few are the ones that face the internet. Microsoft says attacks on public-facing systems rose from 15% to 24% of the intrusions it studied. For a small business that means a short list:
- your firewall or router
- your VPN or remote access tool, the software your IT partner uses to reach your computers
- your website and its plugins
- anything you log in to from outside the office that runs on your own equipment
There is also a new item. Google counted 782 flaws this year in tools for building AI agents, such as Flowise and Langflow, and 212 in tools for running AI models, such as Ollama and LiteLLM. These are often set up by one curious employee on a work laptop or a cloud account, with no owner and no update plan. Google says nearly a quarter of the model-serving flaws involved open connections that needed no login or let an attacker reach other systems.
A free account opens the rest of this article and the whole library.