AI finds flaws
faster than you can patch them
We often hear that artificial intelligence will replace system administrators. In the field, it is the other way round: AI creates work for us. AI finds flaws. Every flaw becomes a CVE, and every CVE becomes an update to apply on your servers.
About 48,000 CVEs were published in 2025, roughly 130 a day. Here is what that changes for whoever runs servers, and why an SMB needs someone to look after them more than ever.
1. AI has become a flaw-finding machine
~40,000
CVEs published in 2024
~48,000
CVEs published in 2025
~130
CVEs a day in 2025
Counts differ by a few hundred depending on the source and snapshot date (The Stack, Zafran).
Vulnerability research has been tooled up. In 2024, Google's Big Sleep agent found an exploitable flaw in SQLite, fixed before it reached a released version. In 2025, XBOW, an autonomous penetration-testing system, climbed to the top of HackerOne's US leaderboard. Vendors and researchers now comb their code with language models.
To be fair, not all of the increase comes from AI. Since February 2024, the Linux kernel has been assigning its own CVEs and publishing several thousand a year. But for whoever runs the servers, the cause hardly matters: there are more patches to assess, and each one takes time.
The flip side: noise
AI also produces fake security reports, plausible and useless. The curl project closed its bug bounty on 31 January 2026 for that reason (The Register). Its maintainer, Daniel Stenberg, also counts more than a hundred good AI-assisted reports. AI produces both: real flaws and noise. Sorting them out is work.
2. The same AI helps exploit flaws
A published patch shows where the flaw was. Reading that patch and turning it into an exploit used to take a specialist days. With a language model, that time shrinks.
Erlang/OTP SSH (CVE-2025-32433), April 2025
A researcher got a working exploit with GPT-4, starting from the patch, in about a day. The first code it produced did not work: the model had to be guided step by step (The Register). AI does not replace the attacker: it makes them faster.
University of Illinois, April 2024
A GPT-4-based agent exploited 87% of a sample of 15 already-published flaws, from the CVE description alone. Without that description, the rate fell to 7% (TechRepublic). The sample is small, but the lesson is clear: once a flaw is described, exploiting it is within reach of far more people.
The window is short: two recent cases
- PHP-CGI (CVE-2024-4577), June 2024. An argument injection in PHP running as CGI on Windows, mass-exploited in the days after it was published.
- React2Shell (CVE-2025-55182), December 2025. Remote code execution in React Server Components and Next.js. Within hours of publication, GitHub filled up with AI-generated proofs of concept, nearly all useless, some booby-trapped with malware (VulnCheck). A working proof of concept came out the next day. Then AI-written malware exploited the flaw to mine cryptocurrency on unpatched servers (SC World).
The number to remember
According to VulnCheck, a quarter to a third of exploited flaws are attacked on the day they are published, or earlier: 28.96% in 2025, 23.43% in the first half of 2026 (VulnCheck). The figure moves from one period to the next, but it stays high: for an internet-facing server, waiting for next month's maintenance window is no longer an option.
3. What it changes for whoever runs the servers
Monthly patching is no longer enough
You need to watch continuously, sort out what really affects your servers and quickly apply what is exposed. The rest can wait for a planned window.
Emergencies land at night and on weekends
A critical CVE does not wait for office hours. Someone has to be available to read the advisory, assess exposure and apply the patch.
Being able to roll back
Patching fast raises the risk of regressions. A recent, verified backup taken before the change lets you patch urgently without risking everything.
AI crawlers load servers too
Another, quieter effect of AI: language-model crawlers consume CPU and bandwidth. Rate limiting and filtering are now part of routine operations.
4. How we triage: 130 CVEs a day, which ones affect you?
Of the dozens of CVEs published every day, only a handful touch any given server. The whole point is to spot them fast. Here is how we go about it.
Routine updates go out on their own
Routine security updates are applied automatically. Human time is kept for what needs it: emergencies, restarts, version upgrades.
An up-to-date inventory, server by server
Every hour we record installed packages and the version available, and the inventory is refreshed right after each update. At any time we know what is running, and in which version.
Cross-checking against CVEs
That inventory is matched against CVEs, package by package. A flaw in software you do not run does not concern you; a flaw in an exposed service goes to the top of the list.
AI as a triage aid
AI agents help us prioritise: summarising a security advisory, spotting CVEs that hit an exposed service, flagging those that require a restart. AI sorts; the decision stays with a human, as does choosing when to restart.
Critical flaw: patched within 48 hours at most
A critical CVE is patched within 48 hours of the vendor releasing the fix. That is a ceiling, not the normal timeframe: an exposed, actively exploited flaw is handled as soon as the fix exists. If the patch requires a service interruption (restarting a database or a hypervisor), we schedule it with you.
5. Why managed services matter more than ever
An SMB does not have someone available to follow 130 CVEs a day, sort out the ones that matter and step in when the patch comes out. That work already existed; AI has made it bigger and more urgent. It is exactly what managed services take on: operational maintenance, monitoring and intervention.
Hiring someone in-house is still possible, but one person covers neither holidays nor nights: we ran the numbers in hire a sysadmin or outsource. And for an emergency patch to be applied on a Sunday night, you need 24/7 on-call support, included in our Pro and Critical plans.
Updates in our plans
Beyond the included hour on Pro and Critical: €120 excl. VAT per hour, counted in 30-minute increments. Whenever possible, you are notified before any overrun; for a critical flaw, intervention does not wait. When your prepaid hours are close to zero, you are notified.
Frequently asked questions
Will AI replace system administrators?
Today, no: it shifts the work and adds to it. In the longer term, nobody can predict. AI helps triage alerts and CVEs, read logs or write a script. But it also creates work: the flaws it finds become patches to apply, test and schedule. The decision to apply a patch, restart a service or warn a customer stays with a human.
How many CVEs are published each year?
About 40,000 in 2024 and about 48,000 in 2025, roughly 130 a day. Counts differ by a few hundred depending on the source and snapshot date. Only a small share affects any given server: the whole point is triage.
How often should security updates be applied?
Routine security updates can be applied automatically every day. Critical flaws on internet-facing services are handled as soon as the patch is out: a share of exploited flaws is attacked on the very day they are published. A monthly cycle is no longer enough for an exposed server.
Are security updates included in RDEM managed services?
On Pro (€150 excl. VAT per month per server) and Critical (€250 excl. VAT), updates are covered by the one hour per server per month included, which covers routine needs. On Essential (€70 excl. VAT), operational maintenance and 24/7 monitoring are included, with no intervention time credit: routine security updates are applied automatically, and time spent intervening is billed at €120 excl. VAT per hour, counted in 30-minute increments.
Are your servers up to date tonight?
Once your servers are with us, we inventory them, match their packages against published CVEs and patch what is exposed first.