Home/ Blog/ Security news/ Article
Signal · Weekly Analysis

Four of this week's flaws would show up first as a crash nobody sees

NoisyTCP, Apache mod_http2, NetScaler and headless Chrome would first show up as crashes that self-healing servers erase. How to keep that crash evidence.

Dark cover with the headline: four flaws would show up first as a crash nobody sees

If someone spent this week trying to turn one of the new memory bugs into a working exploit against your servers, the first trace would not be an alert. It would be a restart. And most of the machinery on a modern server exists to make a restart go unnoticed.

Four of the five vulnerabilities we covered in the week to November 17 share that first symptom. NoisyTCP panics the Linux kernel. The Apache mod_http2 race kills a child process. Citrix describes every one of the 28 NetScaler memory flaws as unpredictable behavior or denial of service. The Chrome XML zero-day also lives in the headless renderers that servers run, where a failed attempt looks like a renderer crash the service quietly recovers from. The fifth, the Apache Axis flaw behind the Confluence and N-central attacks, crashes nothing at all, and that contrast is useful too. The table below sets out where each one's first evidence lands and what removes it.

FlawFirst visible signWhere it is recordedWhat removes it
NoisyTCP, Linux kernel CVE-2026-99288Kernel panic, then the whole host rebootsPrevious boot's kernel log, kdump, hypervisor serial consoleA journal kept only in memory, and the reboot itself
Apache httpd CVE-2026-119618 (mod_http2)A child process dies with a segmentation faultAH00052 line in the Apache error logThe parent starts a replacement child; log rotation
NetScaler CTX697305 (28 memory flaws)Appliance restart or high-availability failoverAppliance event and HA logsThe secondary node takes over and service looks healthy
Chrome CVE-2026-125350 in headless renderersRenderer crash inside a PDF or screenshot serviceThe service's own application log, if it logs crashesJob retry and a fresh browser launch on the next request
Apache Axis CVE-2026-90436 (Confluence, N-central)No crash: a SOAP request, then a new child process of JavaProxy and Tomcat access logs, process recordsNothing removes it, but nothing alerts on it either
Where the first sign of this week's flaws lands, and what quietly removes it. Sources: CVE-2026-99288 record and researcher advisory, Apache httpd security page, Citrix CTX697305, Google and CISA for CVE-2026-125350, Rapid7 and Atlassian for CVE-2026-90436.

Read the table by its last column

The second and third columns are the ones defenders already know. Advisories print crash signatures, and our own posts this week gave the strings to search for: AH00052 for Apache, a NULL pointer dereference in tcp_v4_rcv for NoisyTCP, unexplained restarts and failovers for NetScaler. The last column is where the evidence goes, and nearly every entry in it is a reliability feature doing its job.

Apache's parent process replaces a dead child within seconds, so a crashing worker costs one failed request and the site stays up. A NetScaler high-availability pair hands traffic to the secondary node, which is the whole point of buying two. A rendering service retries the job and launches a fresh browser. A Linux host that panics comes back on its own, and on distributions where the journal lives only in memory, the kernel log of the boot that crashed is gone. Each of these was built so that a crash does not page anyone. That is the right design for uptime. For detection it is a blind spot: the failed attempt leaves its mark in a log that nobody reads, or in one that no longer exists.

A crash is what an exploit attempt looks like before it works

Our first original claim is about what those crashes mean. Turning a memory-corruption bug into code execution is rarely one clean shot. It is repeated attempts at a race or a heap layout, and most of those attempts end with the target falling over. The public tooling for the Apache bug, as we described it on November 12, drives the race until a child process crashes. Hyunwoo Kim, who found NoisyTCP, says in his advisory he has so far written only a limited, lab-specific exploit aimed at virtual machines and is still researching a general one. In his tests a virtual machine panicked within 5 seconds to 2 minutes, while a quiet bare-metal server took about 14 hours.

Put those together and the crash stops being a reliability event. It is the attacker's development log, written on your hardware. A run of crashes on an internet-facing host, with no deploy and no load spike to explain it, is the most likely way you would see someone working toward code execution in the gap between disclosure and your patch. The NetScaler bulletin, CTX697305, makes the point harder to ignore: its two headline entries, CVE-2026-120868 and CVE-2026-107412, carry high confidentiality and integrity ratings in their CVSS vectors even though the text says crash. In our reading, a vendor describing an outcome as denial of service is describing what it has confirmed, not necessarily the ceiling.

The patch can erase the window you most need to see

Our second claim is about timing. Patching is a restart. A kernel fix for NoisyTCP only applies after a reboot. A NetScaler upgrade reboots the appliance. Apache needs a full restart to load 2.4.71. On a host without a persistent journal, the reboot that installs the fix also discards the kernel log from the days the host was exposed, and a log rotation scheduled between disclosure and patch can do the same for Apache's error log.

So the order matters. Collect crash history first, then patch. It costs a few minutes per host, and it is the only way to answer the question every advisory leaves open: did anyone try before we closed it? Once the reboot has happened, the honest answer for many fleets is that they cannot know.

The exploited bug this week did not crash anything

Our third claim is the limit of all this, and the fifth row of the table shows it. The flaw with confirmed attacks this week, CVE-2026-90436 in Apache Axis 1.x, is unsafe reflection, not memory corruption. A crafted SOAP fault makes the server load a class the attacker names, and the application keeps running as if nothing happened. Rapid7 caught the Confluence attack on November 13, and the bug class needs no crash to get there.

That matches the shape of our own ledger. Of the 6,465 vulnerabilities we triaged in the seven days covered by this column, the largest bug class was command injection and remote code execution, at 541 entries, ahead of cross-site scripting at 339. Those classes succeed cleanly: the injected command runs exactly as written. By our tracking, around 43 of the 6,465 carried an exploitation signal in the same window. Crash monitoring catches the memory bugs on their way to working. It will never catch the logic bugs, where the evidence is what happens after: a new child process of the Java service, an outbound connection from a host that should not make one. A detection plan built only on crashes covers one half of this week's risk.

The two halves meet in one place. Our advice for headless Chrome hosts on November 11 was to watch for a renderer crash followed by an unexpected child process. That pairing is the strongest signal on the list: a crash shows someone working on the bug, and a new process shows it worked.

Make crashes survive, then make them count

None of this needs new tooling. It needs the crash evidence to outlive the restart and to reach someone who compares it with a baseline. In order:

  1. Keep the kernel's last words. Set Storage=persistent in /etc/systemd/journald.conf (or create /var/log/journal) so journalctl -k -b -1 still has the previous boot after a panic. Where you can, enable kdump or keep the hypervisor's serial console log, because a panic often never reaches the disk.

  2. Ship crash lines off the host. Forward the strings that mean a process died: Apache's AH00052 with Segmentation fault or Aborted, kernel Oops and NULL pointer dereference, and systemd's code=dumped exits. A crash record held only on the box that crashed is one rotation or reboot away from gone.

  3. Alert on the rate, not the event. Single crashes happen on every busy fleet. What matters is change: a host that logged no child segfaults last month and logs a cluster in an hour, or several hosts rebooting on the same day with the same function in the trace.

  4. Treat a failover as an incident, not a success. On HA pairs, record why the primary dropped before you fail back. For NetScaler, an unexplained failover on an unpatched build is exactly the signal the bulletin's crash language points to.

  5. Harvest before you patch. On each exposed host, run journalctl --list-boots, search current and rotated error logs for AH00052, and export appliance event history before the maintenance reboot.

  6. Join a crash to what follows it. For services that render untrusted input or run Java, alert when a crash is followed within minutes by a new child process or an outbound connection to a destination the service never contacts. That join catches the exploit that finally worked, and the logic bug that never crashed at all.

This week's patches close the holes. The crash logs are how you find out whether anyone reached them first, and only if the restart has not wiped them.

Methodology: figures are drawn from the Suriq threat desk's own intelligence and news ledgers over the stated window. "Triaged" counts events we logged a decision on, not raw signal volume; exploitation figures are best-effort and labelled approximate.

Topics

Frequently asked questions

Why would an exploit attempt against these bugs show up as a crash?

Turning a memory-corruption bug into code execution usually takes repeated attempts at a race or heap layout, and failed attempts crash the target. NoisyTCP panics the Linux kernel, the Apache CVE-2026-119618 tooling drives the race until a child process crashes, and Citrix describes the 28 CTX697305 flaws as causing unpredictable behavior or denial of service.

What erases crash evidence on a typical server?

Reliability features. Apache's parent replaces a dead child within seconds, a NetScaler HA pair fails over to the secondary, rendering services retry with a fresh browser, and a panicked Linux host reboots. On distributions where the journal lives only in memory, the kernel log of the crashed boot is lost.

How do I keep kernel panic evidence for NoisyTCP CVE-2026-99288?

Set Storage=persistent in /etc/systemd/journald.conf or create /var/log/journal, so journalctl -k -b -1 shows the previous boot. Enable kdump or keep the hypervisor serial console log too, because a panic often never reaches the disk. Search for NULL pointer dereference in tcp_v4_rcv or tcp_v6_rcv.

Should I check logs before or after patching?

Before. Kernel, NetScaler and Apache fixes all need a restart or reboot, and on hosts without persistent logs that restart discards the record of the exposure window. List boots with journalctl --list-boots, search current and rotated Apache error logs for AH00052, and export appliance event history first.

How many vulnerabilities did Suriq triage this week?

In the seven days to November 17, 2026, the Suriq Research Desk triaged 6,465 vulnerabilities, logging a decision on each. Around 43 carried an exploitation signal by our tracking. Command injection and remote code execution was the largest bug class at 541, ahead of cross-site scripting at 339.

See what Suriq finds on your own servers.

Agentless checks for uptime and your cloud, a lightweight agent for detection on each server, and one console that ranks and explains what it finds.