Wednesday’s Patch Tuesday cycle put Microsoft in an awkward spot. The company that sells Defender as your endpoint cybersecurity layer pushed emergency fixes for two of its own components, both actively exploited as zero-days, both letting an attacker climb to SYSTEM or knock the agent flat. UnDefend and RedSun aren’t theoretical. Microsoft says they’re already being used in attacks.

Then The Hacker News dropped CVE-2026-46333: a nine-year-old Linux kernel bug that lets an unprivileged local user read sensitive files and execute commands as root across default installs of major distros. Different vendor, different stack, same shape of problem.

If you spent the last decade buying agents to watch your endpoints, this is the week to ask whether those agents have started watching you back.

The Defenders Got Defended Last

Microsoft’s UnDefend and RedSun bugs sit inside the Defender for Endpoint stack itself. One is a privilege escalation. The other is a denial-of-service that takes the agent down without warning. Both require local access, which sounds reassuring until you remember how attackers operate in 2026: they land via stolen tokens, malvertising, or a poisoned package, and then they need exactly one local primitive to finish the job.

A Defender escalation gives them that primitive on a silver platter. The DoS variant is uglier still. An attacker can blind the very tool your SOC depends on for threat detection before doing anything noisy. Your dashboards stay green while the box gets carved up.

This is the part of the security stack nobody wants to think about. Endpoint agents run with the highest privileges your operating system allows. They have to. They hook into the kernel, inspect memory, intercept network traffic, decrypt TLS, write to protected directories. Every one of those capabilities is also a capability an attacker would love to inherit. When the agent itself has a bug, the blast radius is the whole machine.

Nine Years Is A Long Time To Be Wrong

The Linux kernel flaw makes the same point from a different angle. CVE-2026-46333 has been sitting in mainline kernels since roughly 2017. Researchers found it this month. Default installs of Ubuntu, Debian, and Red Hat derivatives are all vulnerable. The CVSS score is a modest 5.5 because the bug needs local access, and local access is exactly what attackers get for free after a single phishing click or a container escape.

What’s striking is the dwell time. Nine years of code review, fuzzing campaigns, static analysis tooling, and academic kernel research, and the bug survived all of it. The Defender flaws were similarly quiet until someone weaponized them. Boring code in privileged components is a perfect hiding spot. Nobody looks because it works. Until it doesn’t.

This is the lesson cyber security teams keep relearning. The components that grant the most authority, the kernel, the EDR agent, the identity broker, the hypervisor, are the ones most worth auditing and least often audited. Vendor reputation does not change the math.

What To Do This Week

Start with the patches. Microsoft’s KB-numbered updates for Defender went live Wednesday. Push them through your normal ring, then verify the agent version on every endpoint, because a DoS-able agent that quietly downgrades itself is still useful to an attacker. For Linux, watch your distro’s advisory channel for CVE-2026-46333 backports and apply them on anything with multi-tenant or untrusted local users: CI runners, jump boxes, dev workstations, classroom labs.

Then go past patches. A few concrete moves worth the calendar time:

• Treat the endpoint agent as a privileged service. Inventory which hosts run which version. Alert on unexpected downgrades, unexpected stops, and tamper-protection failures. If your EDR has a self-protection log stream, ingest it into the SIEM.
• Build a “blind agent” detection. If a host stops reporting heartbeat past your threshold, the incident response page goes out. Attackers exploiting an agent DoS look identical to a host that simply got unplugged. Treat them the same until proven otherwise.
• Hunt for the prerequisites. UnDefend-style escalation needs prior local code execution. Look for unusual child processes from browsers, Outlook, Teams, and signed installers. Cross-reference those signals against the patch state of Defender on the same host.
• Apply defense in depth to local privilege. Application allowlisting, WDAC or AppLocker in audit mode at a minimum, kernel module signing on Linux, sudo logging shipped to your SIEM. Security hardening at the local-user-to-root boundary is exactly the layer these bugs cross.
• Segment by privilege, not just by network. The host running your EDR console, your patch server, your identity sync agent: those are crown-jewel boxes. Treat them like domain controllers.

A working firewall and decent brute-force protection on management interfaces are still table stakes. They are not the conversation anymore. The conversation is what happens after an attacker is already executing code on your endpoint, which in 2026 is most days.

The Cybersecurity Trust Tax Comes Due

A pattern runs through this week’s stories that deserves a name. Defender flaws, kernel flaws, signed-malware services dismantled by Microsoft, AI-discovered vulnerabilities in mature codebases: every layer the industry has been telling buyers to trust has shipped at least one exploitable bug in the last twelve months. The signed code was signed by stolen certs. The endpoint agent has a privilege escalation. The kernel had a nine-year hole.

The honest read is that “trusted component” has become a category boundary. It tells you which pieces of your stack carry the most authority and therefore deserve the most scrutiny, the fastest patch cycles, and the most aggressive monitoring. Threat-protection budgets that load up on more agents without hardening the agents already deployed are buying additional attack surface at premium prices.

Frequently Asked Questions

Should I disable Defender until the patches are verified?

No. The exploitable surface is still smaller with a patched agent than with no agent at all. Push the update through your normal ring, verify the version, and monitor for tamper or downgrade events.

Does the Linux kernel flaw affect containers?

Yes. CVE-2026-46333 is a host kernel issue, so any container sharing that kernel inherits the exposure. Multi-tenant Kubernetes nodes and CI runners with untrusted job content are the highest-risk profiles.

How do I detect a silently disabled EDR agent?

Heartbeat absence is the simplest signal. Pair it with process telemetry from a secondary source (Sysmon on Windows, auditd on Linux) so an attacker can’t blind both with a single primitive.

Sources

• Microsoft warns of new Defender zero-days exploited in attacks
• Microsoft Patches Exploited UnDefend and RedSun Defender Zero-Days
• 9-Year-Old Linux Kernel Flaw Enables Root Command Execution on Major Distros