Patch window shrinks to hours: three critical flaws in five days

Three flaws and one pattern

The week left three critical vulnerabilities with a common trait worth a slow look.

On September 11, Check Point published patches for two remote code execution flaws in its VPN appliances, CVE-2026-85102 and CVE-2026-85103 (Source: securityweek.com).

The first affects the Security Gateway and allows code execution without credentials.

The second also reaches the Management Server.

The Dutch NCSC warned of imminent exploitation, and LivePatch Take 24 covers R81.20, R82, and R82.10, while the rest of the versions require upgrading to R82.20 (Source: blog.sied.ar).

The operational recommendation that comes with the advisory is concrete, since ports 500, 4500, and 443 should be restricted to authorized peers only.

That same day... GitLab shipped patches for CVE-2026-85706, a path traversal in the commits API with a CVSS of 10.0. OMG!

The flaw allows reading files from the server without authentication. OMG!

Probes in production were detected hours after public disclosure, according to watchTowr (Sources: thehackernews.com, elhacker.net).

The fixed versions are 19.3.2, 19.2.6, and 19.1.8.

The third case is N-able N-central, CVE-2026-86218, a pre-auth RCE with a CVSS of 10.0.

Live exploitation entered CISA's KEV catalog with a remediation deadline of September 11, and Shadowserver counted close to 1,500 exposed consoles (Source: thehackernews.com).

The pattern

There is one scene that sums up why these three cases read together.

In the GitLab case, probes showed up in production hours after the patch went public. There was no grace window.

Someone was watching the advisory repository, or had a system watching it for them, and acted before most administrators had read the email.

That is what stands out!

Public disclosure no longer marks the start of a days-long deadline.

It marks the starting gun of a race that some run with automation.

The N-able case adds another layer.

CISA added the flaw to its catalog of exploited vulnerabilities with a one-day remediation deadline.

That deadline is not rhetorical, that is, it reflects that exploitation was already happening when the advisory was published.

The 1,500 exposed consoles Shadowserver counted are the surface where that clock runs.

Not an isolated case

The three cases share a pattern that the previous week had already shown more bluntly.

GreyNoise and SecurityWeek documented a campaign with hundreds of AI agents against PaperCut NG/MF servers, chaining CVE-2026-81578 and CVE-2026-82078 on a Codex harness with a DeepSeek model.

The result was around 440 organizations compromised across 48 countries, almost half of them in the education sector (Source: bleepingcomputer.com, greynoise.io).

One education center went from initial access to Domain Admin in seven minutes.

The connection is not that the flaws are the same, it is rather that the cost of finding, testing, and deploying an exploit has dropped.

A low-budget attacker can rent AI agents to scan, generate variants, and validate the exploit in their own lab before launching it.

What used to require a human team for days now runs in parallel and at scale.

That is why the patch window has shrunk, but not because the flaws are more severe, rather because on the other side there are more hands, faster and cheaper.

When the agent attacks

It is worth saying plainly, even if it stings. I have the feeling that we are losing control on defense ever since agents burst onto the scene.

In the PaperCut case... the AI agents were not the defensive tool, they were the attacker.

Hundreds of instances orchestrated to discover or test and deploy exploits against exposed servers.

The Codex harness with a DeepSeek model is the technical piece that makes it possible without more... an environment that runs tasks autonomously on a language model, but believe me, I have tried simple approaches in my lab with small local models that fit on any GPU, and you can hide behind several proxies, etc... better not to give ideas.

That changes the conversation about agents in the SOC.

It is not just that an analyst can use an agent for triage.

It is that the same kind of technology, with less supervision, is already used to attack.

The asymmetry is real. Unfortunately, the defender needs everything to work and the attacker only needs one unpatched server.

My own read

For anyone administering these systems, the message is uncomfortable but clear.

Patching in hours is no longer the minimum to stay off the victim list.

Segmentation and egress filtering remain the controls that make the difference when exposure time is measured in hours, not days.

And there is an inventory lesson, which is that the 1,500 exposed N-able consoles were not hidden, no... they were on the internet, reachable.

The same goes for the Check Point VPN appliances and the GitLab instances.

The exposed surface is the first problem and, therefore, patching is the second.

What to look at

• Check exposed ports. If you manage Check Point appliances, restrict ports 500, 4500, and 443 to authorized peers. If you have N-able N-central consoles reachable from the internet, that is the first decision on Monday.
• Check GitLab versions. The fixed versions are 19.3.2, 19.2.6, and 19.1.8. If your instance is below that and exposed, the path traversal allows reading files without authentication.
• Enable probe telemetry. GitLab probes were detected hours after disclosure. A log of access attempts to commits API paths gives you the signal before the exploit works.

How I would test it in my lab

I would set up a GitLab instance on an isolated virtual machine, with no internet exposure, and send requests to the commits API with manipulated paths to see whether the server returns file content outside the expected directory.

Before that, I would capture traffic with tcpdump to have a record of which requests arrive and in what shape.

The conclusion would be that, if the vulnerable version responds with file content it should not, the flaw is real and the patch is urgent.

If it responds with an error, the version is already fixed.

It is an afternoon check that requires nothing more than a VM and a willingness to read HTTP responses.

Closing

The window between disclosure and exploitation has compressed to hours.

Anyone without a patching process that works in that timeframe is playing a different game.

The good news is that the controls that make the difference are still the same as always, less exposed surface, fast patches, and telemetry that warns before the exploit arrives.

---

Originally published at https://sammideblas.com/notas/patch-window-shrinks-to-hours-three-critical-flaws-in-five-days